Note: The links to the talks are available under the drop-down menu of each presentation description.
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
8:30-17:00
Monday 21 October 2024 – Day 1
9:00-10:30
Plenary Session I (Room Castello): Edge AI Technology Advancements
9:00-09:15
PIONEERING THE FUTURE OF EDGE AI
Ovidiu Vermesan, SINTEF, Norway.
In today’s fast-evolving technological landscape, edge AI stands at the forefront of innovation, enabling data processing at the edge, closer to the source of information. The presentation delves into the transformative potential of edge AI in enhancing real-time data analysis, reducing latency, and improving privacy. It also introduces the topics of the conference. By leveraging advanced algorithms and machine learning at the edge, we can empower devices to make intelligent decisions, fostering a new era of efficiency, security, and connectivity. Join us as we envision a world shaped by edge AI technology.
Presenter: Ovidiu Vermesan, SINTEF, Norway.
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies and applying the developments to applications such as green mobility, energy, buildings, autonomous systems, electric connected, autonomous, and shared vehicles, and industrial manufacturing. He is currently the coordinator of the Edge AI Technologies for Optimised Performance Embedded Processing (EdgeAI) project.
9:15-10:00
A SUSTAINABLE AI-DRIVEN TRANSFORMATION THANKS TO EDGE AI
Alessandro Cremonesi, Executive Vice President, Chief Innovation Officer General Manager, System Research and Applications STMicroelectronics
In a digitalized and interconnected world, where reality and virtuality blend, AI will be omnipresent from cloud to edge. Distributed AI processing enables sustainable deployment of new services and applications. By enhancing energy efficiency, privacy, and security, Edge AI can ensure scalability to the next level of digital transformation. The semiconductor industry, crucial to this evolution and driven by AI, is expected to reach $1 trillion by 2030. To sustain this growth, the industry will leverage the full range of technologies from ‘Moore’s law’ to ‘More than Moore’ to ‘heterogeneous integration’. New architectures surpassing Von Neumann limitations and algorithms suited for edge constraints will be necessary. AI is driving innovation, demanding new developer profiles that will depend on the semiconductor industry to provide agile and user-friendly tools to harness all technologies effectively.
Presenter: Alessandro Cremonesi, STMicroelectronics, Italy.
Alessandro Cremonesi is STMicroelectronics’ Executive Vice President, Chief Innovation Officer and General Manager of STMicroelectronics’ System Research and Applications (SRA) Group. He has managed the SRA group since 2013 and added the Innovation Office to his mandate in early 2020. Cremonesi’s responsibilities span from global innovation coordination to corporate advanced R&D to system-solutions support for ST customers.
Cremonesi joined STMicroelectronics in 1984. He has served in managerial roles with both Strategic Marketing and R&D responsibilities across domains from telecommunications to audio/video digital-signal processing and multimedia applications. He has been a key contributor to ST’s extensive efforts and strategy in IoT and Artificial Intelligence and, more recently, has led the creation of strategic initiatives to increase ST’s innovation capability.
Cremonesi was part of an expert advisory group defining the agenda for Artificial Intelligence for the Italian Ministry of Economic Development. He has authored several technical papers and patents and is a member of the Scientific Advisory Board at IMEC.
Born in Sant’Angelo Lodigiano, Italy in 1958, Alessandro Cremonesi graduated with a master’s degree in Electronics Engineering from University of Pavia in 1984.
10:00-10:30
Project Pitch Presentations – NEUROKIT2E, dAIEDGE, EdgeAI-TRUST, H2TRAIN, ANDANTE (Room Castello)
10:30-11:00
Morning Coffee/Tea Break (Room Terrazza)
11:00-17:00
Parallel Sessions
11:00-12:30
Session A01 (Room Stampace): EDGE AI TOOLS AND METHODS.
Session Co-chairs:
Mario Diaz Nava, STMicroelectronics, France; Tommie Kerssies, Eindhoven University of Technology, The Netherlands.
The session focuses on the essential frameworks and techniques that empower the development and deployment of edge AI solutions. Speakers will explore various tools for data processing, model training, and inference at the edge, including software libraries, hardware platforms, and edge computing architectures. We will discuss best practices for optimising machine learning models for edge environments and real-world case studies illustrating the successful implementation of these methods. This session aims to equip attendees with the knowledge to harness Edge AI technologies effectively for innovative applications.
Session B01 (Room Castello): EDGE AI MACHINE VISION.
Session Co-chairs:
Alain Pagani, DFKI, Germany; Ana Pinzari, Institute of Engineering Univ. Grenoble Alpes, TIMA, France.
The session delves into integrating edge AI with machine vision technologies, revolutionising how systems interpret and analyse visual data. Participants will explore cutting-edge applications where real-time image processing and object recognition are critical. We will discuss the advantages of deploying AI models at the edge, including reduced latency, enhanced privacy, and improved efficiency. This session aims to highlight the transformative impact of Edge AI on machine vision, showcasing innovative solutions and future trends in the industry.
11:00-11:20
Presentation: A01-1: RECENT TRENDS IN EDGE AI: EFFICIENT DESIGN, TRAINING AND DEPLOYMENT OF MACHINE LEARNING MODELS.
Maen Mallah, Fraunhofer Institute for Integrated Circuits, Germany.
With a rising demand for ubiquitous smart systems, processing and interpreting large quantities of data generated on the edge at a high velocity is becoming an increasingly important challenge. Machine learning (ML) models such as Deep Neural Networks (DNNs) are an essential tool of today’s artificial intelligence due to their ability to make accurate predictions given complex tasks and environments. However, Deep Learning is computationally complex and energy intensive. This seems to contradict the characteristics of many edge devices, which have only limited memory, computational resources, and energy budget available. To overcome this challenge, an efficient ML model design is crucial that incorporates available optimization techniques from hardware, software, and methodological perspective to enable energy-efficient deployment and operation on the edge. This work comprehensively summarizes recent techniques for training, optimizing, and deploying ML models targeting edge devices. We discuss different strategies for finding deployable ML models, scalable DNN architectures, neural architecture search, and multi-objective optimization approaches, to enable feasible trade-offs considering available resources and latency. Furthermore, we give insight into DNN compression methods such as quantization and pruning. We conclude by investigating different forms of cascaded processing, from simple multi-level approaches to highly branched compute graphs and early-exit DNNs.
Speaker: Maen Mallah, Fraunhofer Institute for Integrated Circuits, Germany.
Maen Mallah is a researcher in the area of embedded AI at the Fraunhofer Institute for Integrated Circuits (IIS). He obtained his B.Sc in Telecommunication Engineering in 2014 from An-Najah National University, Palestine and his M.Sc in Electrical Engineering in 2018 from Bilkent University, Turkey with a thesis titled “Multiplication Free Neural Networks”. In March 2018, He joined Fraunhofer IIS as an expert for eAI and focusing mainly on energy efficient NNs. His main work and interest focuses on implementing and optimizing NNs for Edge applications and designing the SW tools required for such a task with a special focus on Quantization- and Fault-aware training.
Presentation: B01-1: OPTIMIZED TRAINING FOR DEFECT CLASSIFICATION.
Angelo Genovese, Università degli Studi di Milano, Italy.
The presentation addresses the latest developments in optimised training for defect classification developed in the EdgeAI project.
Speaker: Angelo Genovese, Università degli Studi di Milano, Italy.
Angelo Genovese received the Ph.D. degree in computer Science from the Università degli Studi di Milano, Italy, in 2014. He has been an Associate Professor in Computer Science with the Università degli Studi di Milano since 2022. He has been a Visiting Researcher with the University of Toronto, Toronto, ON, Canada. Original results have been published in over 60 papers in international journals, proceedings of international conferences, books, and book chapters. His current research interests include signal and image processing, artificial intelligence, industrial and environmental monitoring systems.
Presentation Link
11:20-11:40
Presentation: A01-2: AIDGE-GUI: A GRAPHICAL USER INTERFACE FOR AI ON THE EDGE.
Fabian Chersi, CEA-LIST, France.
In the last decade, Deep Neural Networks (DNNs) have seen and exponential adoption in a multitude of technical and non-technical fields. As their performances and capabilities have grown, so have their complexity and difficulty in customization and deployment. Some steps forward in this direction have been taken, for example proposing model exchange formats, such as ONNX, that are flexible and widely available, however if a model needs to be adapted or optimized for a specific use case, data set, hardware constraint, or device, this requires a considerable programming and tool knowledge. To alleviate this problem here we present Aidge-Gui, a graphical user interface for visualizing, optimizing and deploying DNN models. This platform-independent, low-weight graphical interface allows to simply load DNN models in different formats (ONNX, JSON, PTH), visualize them in an innovative way to understand their structure and key points, obtain detailed information about layers, edit layer meta-parameters, perform simplifications such as layer fusion, quantize layers at arbitrary precision levels, train networks and finally export them in C/C++ for deployment and compilation on hardware such as CPUs, GPUs, but also low-power embedded devices such as STM32 but as well as custom architectures.
Speaker: Fabian Chersi, CEA-LIST, France.
Fabian Chersi, CEA-LIST, France.
Fabian Chersi has a BSC and a Master in Electronic Physics, and a PhD in Computational Neuroscience and Robotics. After his studies, he initially did a postdoc at the Institute for Neuroinformatics in Zurich and at UCL in London, then he moved to Paris where he worked in private companies on the development of neuromorphic algorithms and ASICs for deep neural networks. He currently works at the Alternative Energies Commission (CEA) in Paris as a senior AI architect on low power AI chips for the edge.
Presentation Link
Presentation: B01-2: DISCOVERING AND CLASSIFYING DIGITAL AND WOODEN INDUSTRIES PRODUCTS’ DEFECTS AT THE EDGE BY A YOLO/RESNET-BASED APPROACH AND BEYOND.
Alessandro Strano, Deepsensing srl, Catania, Italy.
Presents an AI-based AOI software for both digital and wooden industries developed within the EdgeAI project. Current approaches rely on centralized solutions, where the computation is performed inside the inspection machine itself. Instead, we present algorithms that work at the edge to give rise to competitive solutions to existing ones. In particular, we experiment with three different tasks of defect identification: detecting the defect position within a wooden panel by using YOLO, in which a 96% accuracy is reached. Secondly, we perform classification between defective and non-defective wooden panels by using YOLO, reaching a 90.3% accuracy. Finally, concerning the digital industry, we perform a two-step classification between defective and non-defective microchips and then between four possible defect classes in their surface exploiting a ResNet network obtaining a 97% accuracy. We also exploit explainability tools to understand which parts of the images caused the model’s decision. After developing the AI models, we port them to two less power-consuming edge devices, Nvidia OrinNano, and Nvidia OrinAGX, observing unchanged performance.
Speaker: Alessandro Strano and Robin Faro, Deepsensing srl, Catania, Italy.
Alessandro Strano is a computer engineering graduate currently pursuing a master’s degree in the same field. He works for DeepSensing S.R.L. as a research collaborator.
11:40-12:00
Presentation: A01-3: A NEW SOFTWARE APPROACH TO EXECUTE AI/ML INFERENCE FOR HIGH-RELIABILITY APPLICATIONS ON SPACE QUALIFIABLE HARDWARE.
Pablo Ghiglino, Klepsydra Technologies, Switzerland.
The next generation of spacecraft must achieve higher autonomy to handle tasks like space exploration, in-orbit servicing, and space debris management, requiring advanced sensors and increased computational power. Recent advancements in imagery sensor technology and powerful processors are significantly enhancing data capture and processing capabilities. Future spacecraft will generate data at unprecedented rates, necessitating versatile onboard software that can adapt to diverse mission requirements. Adherence to ECSS standards is crucial for ensuring the reliability and safety of space systems, especially for AI algorithms tailored to space environments. Building on Klepsydra’s previous AI work, supported by ESA, this research advances AI for flight software by porting innovative inference engine software to space-qualified hardware, including the GR740, GR765, and RISC-V, using the ECSS-compliant RTEMS6 real-time operating system. The KATESU initiative demonstrated significant improvements in data processing rates and power efficiency, paving the way for genuine AI in space. Reviews the ongoing development of Klepsydra AI inference, focusing on the integration of hardware accelerators like Ramon.Space’s RC64, FPGAs, and GPUs. These developments mark a new era in onboard data processing, offering groundbreaking improvements in performance, flexibility, and integration, driving significant advancements in space-based computing.
Speaker:Pablo Ghiglino, Klepsydra Technologies, Zurich, Switzerland
Pablo Ghiglino has over 10-year experience in the finance industry (e.g. at Bank of America, Credit-Suisse and HSBC in London, Julius Baer in Zurich) as a software developer. This experience exposed him to this high-intensity and high-speed data-processing environment in which data loss (due to losses or delay) results in high economic impact. Parallel to his work in finance, Pablo successfully studied a PhD at the Surrey Space Center in Aerospace Engineering in the field of “High Precision Planetary Landing”. After his studies, he got involved in several navigation projects (automatic drones, spacecraft landing control), where he seized the opportunity and created Klepsydra to respond to the needs for cutting embedded computing by implementing successful software techniques from the highly competitive financial sector in embedded software frameworks for its use in other industries.
Presentation: B01-3: MULTI-STEP OBJECT RE-IDENTIFICATION ON EDGE DEVICES: A PIPELINE FOR VEHICLE RE-IDENTIFICATION.
Tomass Zutis, Institute of Electronics and Computer Science, Latvia.
A method that leverages a multi-step process focused on extracting and using object features for object re-identification is described. The proposed pipeline includes steps such as: detecting an object, converting its features into a vector embedding, storing this embedding in a vector database, and then querying the database to find the same or similar objects based on their feature embeddings. This approach allows for the identification of the same object across different images or cameras, even in varying locations, such as in Vehicle Re-Identification scenarios. For situations where re-identification needs to happen in outdoor environments or on-the-go, implementing this process on edge devices becomes crucial. Therefore, multiple ways to tailor the pipeline and its outputs for edge devices are outlined and investigated. The presentation provides a detailed explanation of the pipeline’s structure and how it functions on edge devices, along with the experimental setup demonstrating its application, particularly in vehicle re-identification.
Speaker: Tomass Zutis, Institute of Electronics and Computer Science, Latvia.
Tomass Zutis is a Computer science bachelor’s graduate and is currently studying for his master’s degree in computer science at the University of Latvia. Tomass has been working at the Institute of Electronics and Computer science since 2023 and is focusing on the fields of Computer vision and Machine learning. Tomass’s various interests in the field include Artificial intelligence, Neural networks, Inference of data, Data analysis and processing and most importantly – Software engineering itself.
12:00-12:20
Presentation: A01-4: EDGE AI FOR BATTERY MANAGEMENT: EFFICIENT SOC AND SOH PREDICTION FOR LI-ION BATTERIES WITH NEURAL NETWORKS.
Silvester Sabathiel, Silicon Austria Labs, Austria.
As battery-driven technologies become more widespread, accurate State of Charge (SoC) and State of Health (SoH) predictions are crucial for effective battery management and predictive maintenance, particularly in resource-constrained edge environments. Traditional methods often rely on computationally expensive processes that are not suitable for deployment on edge devices. This paper presents a novel approach using a single neural network for simultaneous SoC and SoH prediction based on Electrochemical Impedance Spectroscopy (EIS) measurement data and temperature inputs. We conduct a systematic study of neural network architectures, sizes, and quantization techniques to identify the most relevant parameters for efficient operation. The optimized neural network achieves with comparable high accuracy while running on edge devices with a significantly reduced memory footprint and resources, demonstrating the feasibility of advanced battery monitoring in edge AI applications. Our results highlight the potential of deploying sophisticated predictive models on low-resource edge hardware, paving the way for enhanced battery management systems.
Speaker:Silvester Sabathiel, Silicon Austria Labs, Austria.
Silvester Sabathiel earned his Master’s degree in Physics from Karl-Franzens-University of Graz, Austria, before pursuing a Ph.D. in Computer Science at NTNU Trondheim, Norway. His doctoral research focused on neural network models of mathematical cognition, exploring the internal representations and requirements for neural networks to perform mathematical tasks in multi-modal interactive environments. Since 2023, he has been with Silicon Austria Labs, where he primarily works on Physics-informed neural networks and edge-based neural network solutions.
Presentation: B01-4: PRESENTATION.
12:30-13:30
Networking Lunch (Room Terrazza)
13:30-15:30
Session C01 (Room Stampace): RISC-V-BASED EDGE AI HARDWARE ACCELERATION.
Session Co-chairs:
Paolo Meloni, Università degli Studi di Cagliari, Italy; Eike Schultz, University of Lübeck, Germany.
The session focuses on the emerging role of RISC-V architecture in enhancing Edge AI performance through hardware acceleration. Participants will learn how RISC-V’s open-source design enables customised processing solutions tailored for AI workloads at the edge, offering greater flexibility and efficiency. The session explores the advantages of RISC-V in terms of scalability, power consumption, and integration with various sensors. This session will showcase successful implementations and prospects, illustrating how RISC-V is paving the way for innovative edge AI hardware solutions.
Session D01 (Room Castello): EDGE AI ACCELERATORS.
Session Co-chairs:
Hana Krichene, CEA-LIST, France; Peet van Tooren, Almende, The Netherlands.
The session focuses on innovative hardware solutions designed to enhance the performance of edge AI applications. Speakers will explore various types of accelerators, including GPUs, FPGAs, and specialised AI chips, and their role in enabling real-time data processing and analytics at the edge. Key factors, such as power efficiency, scalability, and latency reduction, will be discussed, showcasing successful case studies. This session aims to provide insights into selecting and implementing the right accelerators for optimal edge AI performance.
13:30-13:50
Presentation: C01-1: A RECONFIGURABLE-BASED HETEROGENEOUS AND HIGHLY PARALLEL PROCESSING PLATFORM FOR SAFE AND SECURE AI.
Konstantinos Georgopoulos, Technical University of Crete and National Technical University of Athens, Greece; Iakovos Mavroidis, Technical University of Crete and Exascale Performance Systems (EXAPSYS) plc., Greece.
The integration of AI/ML components into computing solutions is advancing rapidly while unlocking capabilities that were previously unattainable. However, this progress pushes the performance demands on computing devices, often necessitating the offloading of all but the simplest AI tasks onto cloud-based processing. In critical domains such as automotive and healthcare, relying on the cloud is often unreliable or impractical, driving the need for AI computation directly at the edge. To address this challenge, we are developing an innovative System-on-Chip (SoC) designed for edge AI applications. This solution incorporates a CVA6 RISC-V-based processor subsystem, which utilizes 512 MB of HyperRAM as its main memory and an SD card for storage. The processor subsystem is tightly connected to advanced AI and security accelerators, enabling direct memory sharing between the accelerators and the processor. This architecture is designed to deliver an edge AI solution with significantly improved performance, energy efficiency, and security compared to existing systems. Additionally, the support for reconfigurable hardware allows the system to adapt to specific application requirements, extending its functionality without compromising energy efficiency. Alongside these hardware advancements, we will also develop the necessary software infrastructure, including operating systems, hypervisors, and libraries, to fully exploit the hardware’s capabilities.
Speaker: Konstantinos Georgopoulos, Technical University of Crete and National Technical University of Athens, Greece; and Iakovos Mavroidis, Technical University of Crete and Exascale Performance Systems (EXAPSYS) plc., Greece.
Konstantinos Georgopoulos is an affiliated researcher at the Microprocessor and Hardware Lab at the Technical University of Crete and at the Institute of Communication and Computer Systems (ICCS) at the National Technical University of Athens. He received his degree and PhD in Electronic Engineering in 2001 and 2007, respectively, from the Department of Engineering at Lancaster University in the UK. Dr. Georgopoulos has taught and conducted research at both national and European levels, in collaboration with Lancaster University, the Athens University of Economics and Business (AUEB), the Technical University of Crete, and the Foundation for Research and Technology – Hellas (FORTH) for the past 15 years. His primary responsibility has been the technical implementation of complex subsystems based on reconfigurable logic. Some of the most important projects include AFORMI, ECOSCALE, CYRENE, ENERMAN, and SECOPERA. Currently, Georgopoulos holds the management responsibility for two European research projects, REBECCA and MEDIATE, as Project Manager. Additionally, he is a member of the organizing committee of the Key Digital Technologies (KDT) EDGE AI Cluster, an EU initiative aimed at collaboration and coordination of all research achievements and activities in state-of-the-art technologies at the edge utilizing artificial intelligence techniques.
Iakovos Mavroidis is a member at the Technical University of Crete and co-founder/CTO of Exascale Performance Systems (EXAPSYS) plc. From 1991 to 1995, he was awarded two gold and two bronze medals in national competitions and two honourable mentions in International Olympiads in Mathematics and Computer Science. He received his B.Sc. in Computer Science from the University of Crete in 1999 (GPA 9.29/10) and his M.Sc. from UC Berkeley in 2001 (GPA 3.93/4). He has worked at Sun Microsystems, MIPS Technologies—where he led the R20K microprocessor design—Ellemedia Technologies, Virtual Trip, and Kaleao Ltd. From 2004 to 2006, he was also a Visiting Professor at the University of Crete. He received his Ph.D. from the Technical University of Crete in 2011. He has coordinated several EU projects, including RAPID H2020 and OPTIMA EuroHPC, and is currently leading the REBECCA KDT JU project.
Presentation Link
Presentation: D01-1: PLATFORM-INDEPENDENT FPGA AI MAPPING TOOL BASED ON FINN.
Remo Senekowitsch, Zurich University of Applied Sciences (ZHAW), Switzerland.
The FINN machine learning framework is a well-established tool that creates custom AI accelerators using a dataflow-style architecture specifically designed for FPGAs. Currently, the tool is only compatible with a limited selection of AMD FPGAs, and it is designed to generate custom HLS code at a high abstraction level. However, the open-source core of FINN can be utilized for any FPGA. We are developing a new code generator for FINN that directly produces FPGA-RTL code (VHDL or Verilog). This allows for better control over available FPGA resources and enables fine-tuned optimization of the generated code for various aspects such as high throughput, calculation speed, resource usage, processing latency, and more. Our project aims to create a platform-independent AI mapping tool that generates optimized FPGA-RTL code. We are sharing our progress of our work, including the necessary steps, the challenges we encountered, the tools we developed, and we present preliminary results.
Speaker: Remo Senekowitsch, Zurich University of Applied Sciences (ZHAW), Zurich, Switzerland.
Remo Senekowitsch works as a research assistant at the ZHAW Institute of Embedded Systems while studying for his master’s degree in computer science. He is passionate about the Rust programming language and building custom keyboards.
13:50-14:10
Presentation: C01-2: AUTOMATIC MODEL TRAINING, COMPRESSION, AND QUANTISATION TOOLS FOR RISC-V ACCELERATORS.
Timothy Llewellynn, Bonseyes, Switzerland.
Presents a novel suite of automatic model training, compression, and quantization tools tailored for RISC-V accelerators, targeting edge AI applications. As AI models become increasingly complex, deploying them on resource-constrained edge devices with minimal engineering effort remains a significant challenge. Our proposed tool automates the training, optimization and deployment process, leveraging advanced one-shot structed pruning, post-training quantization, and weight distribution techniques to reduce model size and computational requirements without sacrificing accuracy. This work pushes the boundaries of automatic machine learning for edge AI, providing a powerful end-to-end tool for application developers working with RISC-V accelerators. The primary innovation of this work lies in its end-to-end automation of training, compression and quantisation and its tight integration with RISC-V’s inference engines, enabling seamless training and deployment of optimized models in a single step. We demonstrate that state-of-the-art multi-task learning methods to resolve multi-task learning objectives can be applied automatically to joint training and quantisation objectives to produce compressed models with minimal accuracy loss. Experimental results demonstrate substantial improvements in model size and inference speed, with compression ratios up to 20x in memory footprint and acceleration in inference time on Risc-V processors.
Speaker: Timothy Llewellynn, Bonseyes, Switzerland.
Timothy Llewellynn is president of the Bonseyes in Switzerland, which is a non-profit organization focused on democratizing artificial intelligence through collaborative and open AI development, aiming to transform the cloud-centric development of AI into a more decentralized, edge-device-based approach. With over 20 years of experience in high-tech startups and corporate R&D in data-driven software development he holds a B.E. and M.E. with Honours in Electrical Engineering from the University of Canterbury, New Zealand.
Presentation: D01-2: SPIKING AND ARTIFICIAL NEURAL NETWORK ACCELERATORS IMPLEMENTED IN ANDANTE.
Mario Diaz Nava, STMicroelectronics, France.
ANDANTE aimed to implement highly efficient integrated circuits for the AI Edge deep hardware devices. Eight different integrated circuit accelerators were developed and fabricated based on spiking and artificial neural networks. Moreover, six other designs were made in FPGs. The resulting accelerators were validated and evaluated across use cases in 5 application domains. The performances obtained range from micro, milli and tens of Watts depending on the architecture, technology, and associated application. The results show the benefit of using this type of accelerators for highly efficient Edge AI solutions where accuracy, latency and energy consumption are important requirements to meet application needs.
Speaker: Mario Diaz Nava, STMicroelectronics, France.
Mario Diaz Nava has a B.S. in Communications and Electronics Engineering from the Instituto Politecnico National, Mexico. MS. and a doctorate, both in computer science, from the Institut National Polytechnique de Grenoble, France. He has worked at ST since 1990. He has held different positions (Designer, Architect, Design manager, Project Leader, Program Manager) in different ST R&D organizations. His selected project experience is related to specifications and design of communications circuits and management of cooperative programs. He currently holds the position of Manager of Cooperative R&D Programs at ST Grenoble. He is a member of the SRIA Leadership Group. He recently coordinated the ECSEL/ANDANTE project.
14:10-14:30
Presentation: C01-3: TRISTAN – EMBRYO OF A EUROPEAN RISC-V ROADMAP.
Patrick Pype, NXP Semiconductors, Belgium.
Europe needs an open-source ecosystem to drive competitiveness and enable greater and more agile innovation. The use of open-source hardware and software drastically lowers the barrier to designing innovative integrated circuits. The EU-funded TRISTAN project aims to further expand and develop RISC-V architecture in Europe so that is able to compete with existing commercial alternatives. This open specification eliminates the need to learn and create unique ecosystems for each processor architecture, increasing productivity, security and transparency. The TRISTAN approach is holistic, covering both hardware IP developments, as well as electronic design automation tools and full software stacks, and these developments will be integrated in a large number of demonstrators. A number of TRISTAN partners have created a Working Group which has established a European Roadmap for further RISC-V based processor and accelerator developments.
Speaker: Patrick Pype, NXP Semiconductors, Belgium.
Patrick Pype is Director of Strategic Partnerships at NXP Semiconductors and has held this position since 2012. He began his career at IMEC in Belgium in 1986. In 1996 he started the start-up company CoWare on hardware/software co-design which was later acquired by Synopsys. Thereafter he worked at Philips in different positions.
Patrick is Chairman of the AENEAS Technical Expert Group since 2018, and he is member of the INSIDE Steering Board since 2015. He was co-chair of the overall ECSEL Strategic Research Agenda 2019-2021 and is also co-chair of the chapter on “Transport & Mobility”. He currently is chair of a European RISC-V Working Group which defines a European RISC-V strategy and roadmap. Patrick was Chairman of the Engineering Alumni Association of the KULeuven from 2005-2008.
Patrick has contributed several papers in two books published by Springer on “Automated Driving’’ and “Internet of Vehicles’’. He is author of the book “Zest for Opera – Unleash your Leadership”, which was published in October 2018. The book offers inspiring insights from the world of opera for leaders in business & politics, coaches, teachers and students.
Presentation: D01-3: FPGA ACCELERATION FOR SAR IMAGERY-BASED SEA ICE MAPPING.
Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
Accurate sea ice mapping plays an important role in ensuring safe maritime navigation and improving low risk decision-making in icy waters. The Sentinel 1 satellite, which leverages advanced synthetic aperture radar (SAR) technology, provides high-resolution imagery that is useful for detailed real-time monitoring and analysis of sea ice conditions, even under challenging weather and lighting conditions. On-board processing offers a promising solution to the challenges posed by limited downlink capacity and latency in generating final data products. This paper introduces a multi-task U-Net architecture designed to predict sea ice concentration (SIC), stage of development (SOD), and floe size (FLOE) from SAR imagery, utilizing the AI4Arctic dataset. The network is quantized to reduce computational complexity while maintaining performance, thereby optimizing hardware deployment. High-Level Synthesis (HLS) is employed to synthesize the model for efficient implementation on a Field Programmable Gate Array (FPGA).
Speaker: Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
Rashed Al Koutayni, after receiving his master degree in Embedded Computing Systems at the Technical University of Kaiserlslautern-Landau (RPTU), he joined the Augmented Vision department of the German Research Center for Artificial Intelligence (DFKI) in Kaiserslautern (Germany) as a scientific researcher while pursuing his PhD study at the RPTU university. His PhD topic is Hardware Acceleration for Deep Learning Based Computer Vision.
14:30-14:50
Presentation: C01-4: SYNTZULU: A TINY RISC-V-CONTROLLED SNN PROCESSOR FOR REAL-TIME SENSOR DATA ANALYSIS ON LOW-POWER FPGAs.
Gianluca Leone, Università degli Studi di Cagliari, Italy
Spiking Neural Networks (SNNs) are energy- and performance-efficient tools that have been found to be very useful in AI applications at the edge. This talk introduces SYNtzulu, an SNN processing element designed to be used in low-cost and low-power FPGA devices for near-sensor data analysis. The system is equipped with a RISC-V subsystem responsible for controlling the input/output and setting runtime parameters, thus increasing its flexibility. We evaluated the system, which was implemented on a Lattice iCE40UP5K FPGA, in various use cases employing SNNs with accuracy comparable to the state-of-the-art. SYNtzulu dissipates a maximum power of 12 mW when performing SNN inference, which can be reduced to an average of just 1.45 mW through the use of dynamic power management.
Speaker: Gianluca Leone, Università degli Studi di Cagliari, Italy.
Gianluca Leone is an assistant professor at the University of Cagliari, Italy. His research focuses on the design and optimization of FPGA-based systems for signal processing and real-time, edge-deployed SNN-type workloads.
Presentation Link
Presentation: D01-4: PRESENTATION
15:00-15:30
Afternoon Coffee/Tea Break (Room Terrazza)
15:30-17:00
Session E01 (Room Stampace): BENCHMARKING EDGE AI MODELS.
Session Co-chairs:
Gerardo Daalderop, NXP Semiconductors B.V.; Mohammed Abuteir, TTTech, Austria.
The session addresses the critical need for effective evaluation methods to compare and assess the performance of edge AI solutions. Speakers will address the existing benchmarking frameworks, focusing on accuracy, latency, resource consumption, and scalability metrics. They will also address the importance of standardised benchmarks to ensure model reliability across various edge devices and applications. By presenting case studies and best practices, this session aims to empower attendees with the tools and knowledge to assess and optimise Edge AI models for real-world deployment rigorously.
Session F01 (Room Castello): OPTIMISATION METHODS AND TECHNIQUES FOR NEURAL NETWORKS.
Session Co-chairs:
Mario Diaz Nava, STMicroelectronics, France; Pavel Smrz, Brno University of Technology.
The session explores advanced strategies to enhance neural network performance and efficiency. The speakers will address various optimisation techniques, including gradient descent variations, adaptive learning rates, and pruning methods that reduce model size without sacrificing accuracy. The role of regularisation, hyperparameter tuning, and transfer learning in optimising neural networks for diverse applications will be on the agenda. This session aims to equip attendees with practical skills and knowledge to effectively apply these optimisation methods in their projects, driving improved results in machine learning tasks.
15:30-15:50
Presentation: E01-1: CHALLENGES AND PERFORMANCE OF SLAM ALGORITHMS ON RESOURCE-CONSTRAINED DEVICE.
Calvin Galagain, French Alternative Energies and Atomic Energy Commission (CEA); Palaiseau, France, ENSTA Paris, Institut Polytechnique de Paris, Palaiseau, France.
Assessing the performance of Simultaneous Localization and Mapping (SLAM) algorithms is critical for advancing robotic systems. However, the wide variety of hardware options, algorithm settings, and datasets makes it challenging to evaluate SLAM systems comprehensively against the latest advancements. This study aims to test SLAM algorithms on edge devices such as NVIDIA Jetson AGX Orin 64GB, focusing on how these systems perform in real-world scenarios with constrained computational resources. We conduct mapping experiments with various popular visual-based SLAM algorithms that either leverage deep learning models for specific tasks within the process or are learned end-to-end to estimate camera pose. For evaluation, we select RDS-SLAM, which enhances ORB-SLAM3 with semantic segmentation; VDO-SLAM that exploits semantic information to enable accurate motion estimation and tracking of dynamic rigid objects in the scene, TSformer-VO, an end-to-end Transformer-based model for monocular visual odometry; DeepVO, which employs recurrent networks for visual odometry. We also investigate the difficulties in deploying on edge Neural Radiance Fields (NeRF)-based SLAM. We assess these systems using several metrics, including ATE and RPE to measure pose accuracy and rotational drift respectively, as well as runtime, energy consumption, and network latency to evaluate their efficiency and practicality in real-world applications.
Speaker: Calvin Galagain, French Alternative Energies and Atomic Energy Commission (CEA), France.
Calvin Galagain is a first-year PhD student working jointly with CEA and ENSTA Paris on semantic SLAM systems in embedded contexts. Before starting his PhD, he completed his engineering studies at ESIEE Paris, a French engineering school, and the MVA (Mathematics, Vision, Learning) Master’s program at ENS Paris-Saclay. He is passionate about all the challenges of AI and its applications in robotics and various fields of research today.
Presentation: F01-1: ENHANCING SUSTAINABILITY IN WATER RESOURCE MANAGEMENT: OPTIMIZING NEURAL NETWORK ARCHITECTURES FOR ACCURATE PREDICTION OF WATER STRESS IN EUROPEAN COUNTRIES.
Laura Sanz-Martín, University of Salamanca, Spain, and Manal Jammal, IoT Digital Innovation Hub, Spain.
Sustainability in water resource management is critical, given the necessity to monitor and predict key indicators such as SDG 6.4.2: Water Stress. This indicator, which measures the ratio of water resources abstracted relative to their availability, is vital for assessing the pressure on water resources and ensuring their long-term sustainability. The objective of this study is to compare various neural network architectures utilizing different optimisers to predict water stress. The analysis was based on a dataset sourced from AQUASTAT, encompassing data from 28 European countries. Several neural network architectures, with configurations ranging from two to four layers, were implemented and evaluated using optimisers including SGD, Adam, RMSprop, and Adagrad. The findings revealed that a three-layer architecture combined with the Adam optimiser delivered the best performance, achieving an MSE of 0.02187 and an R² of 0.9745, indicating high predictive accuracy. Nevertheless, a two-layer architecture with the SGD optimiser also exhibited strong performance, highlighting its simplicity and effectiveness. These results underscore the importance of meticulous selection of both architecture and optimiser when predicting critical indicators such as water stress. This study not only enhances the accuracy of water-related risk predictions but also supports informed decision-making for sustainable water resource management in Europe.
Speaker: Manal Jammal, IoT Digital Innovation Hub, Spain, and Laura Sanz-Martín, University of Salamanca, Spain.
Manal Jammal is a researcher focused on sustainability challenges, with a background in industrial engineering. She is currently completing her Master’s in Intelligent Systems at the University of Salamanca (USAL), working on leveraging artificial intelligence and innovative technologies for sustainable development.
Laura Sanz is a researcher pursuing a Master’s in Multivariate Data Analysis and Big Data. With a degree in Statistics from the University of Salamanca, her work focuses on applying statistical methods, econometrics, and AI to analyze large datasets. Laura is particularly dedicated to using these tools to address sustainability challenges.
15:50-16:10
Presentation: E01-2: ONLINE AI BENCHMARKING ON REMOTE BOARD FARMS.
Maïck Huguenin, Applied Science University Western Switzerland, Switzerland.
Benchmarks play a crucial role in balancing the benefits and risks of AI by offering quantitative tools that guide responsible development. They deliver objective and consistent metrics for accuracy, speed, and efficiency, enabling engineers develop reliable products and services. Additionally, benchmarks help researchers gain new insights that can drive future innovations. Today, numerous cloud-based AI development services allow software developers to utilize AI models through APIs, SDKs, or applications. Benchmarking these models on cloud infrastructure is a feature offered by these services. However, few of these services are designed for Edge deployment, where deep expertise in embedded programming and system integration is necessary to optimize and deploy AI models on specific embedded devices. Comparing benchmarking results across different embedded boards becomes increasingly complex when targeting devices from various providers. The current project aims to design and implement a collaborative platform that enables researchers and developers to conduct experiments and research across various edge AI domains and edge AI devices. This will be achieved by sharing resources on a distributed platform providing access to dedicated resources, tools, and services, allowing end users without expertise in embedded programming to perform live AI experiments, such as benchmarking, on remote embedded boards.
Speaker: Maïck Huguenin, Applied Science University Western Switzerland, Switzerland.
Maïck Huguenin, an experienced scientist at HES-SO University of Applied Sciences, holds a master’s degree in information and communication technologies with a strong background in low-power electronics, machine learning, and edge AI. He has 10 years of industrial experience in precision manufacturing for mechatronics and healthcare. At HES-SO, his major contributions so far are in the Edge AI domain. He has actively participated to the BonsAPPs project, and he is currently developing the Virtual Lab within the frame of the dAIEdge Network of Excellence. These research activities have contributed to enrich his skills and expertise in the Edge AI field of research.
Presentation: F01-2: PRESENTATION
16:10-16:30
Presentation: E01-3 : PCA-BASED CONFIDENTIALITY PRESERVING METHOD FOR REAL-TIME SOFT-FAILURE DETECTION IN OPTICAL NETWORKS.
Azarm Yeganehfallah, Scuola Superiore Sant’Anna, Italy.
The proposed method assures data confidentiality through permutation and Principal Component reduction of the monitored data. Failure detection leverages clustering algorithms to differentiate optical network states. Results show that the proposed method performs well for small data size, that is suitable for real-time failure detection.
Speaker: Azarm Yeganehfallah, Scuola Superiore Sant’Anna, Italy.
Azarm Yeganehfallah, with extensive industrial experience at a leading telecom company in Iran, she transitioned to academia in 2023 as a research fellow at the prestigious Sant’Anna School of Advanced Studies. Currently, she is pursuing a PhD in the same institute, focusing on her research interests in Software-Defined Networking (SDN), Artificial Intelligence (AI), and confidentiality preservation within the telecom research domain.
Presentation: F01-3: MARKET PATHWAYS FOR CLOUD EDGE IOT IN KEY SECTORS TRANSPORTATION, AGRICULTURE, ENERGY AND MANUFACTURING.
Elisabeth Steimetz and Inessa Seifert, VDI/VDE-IT GmbH, Germany.
UNLOCK-CEI’s (Unlocking the Cloud Edge IoT demand potential in Europe) is a Coordination and Supporting Action funded by the European Commission that aims at accelerating the deployment of the Cloud-to-Edge-IoT (CEI) computing continuum in Europe by focusing on the demand-side drivers and challenges to identify technology-driven innovation and business opportunities driving demand value chains. It carried out a systematic assessment of the state of the open European CEI ecosystem and collected insights on possible business opportunities and market pathways for the European companies. The presentation will cover insights about the innovative solutions including Edge AI applications and Cloud-Edge-IoT technologies that are already available on the market. It will highlight, how these solutions can enable the digital transformation of the key sectors of the European economy such as agriculture, energy, transportation and manufacturing. Participants at the conference will get an overview of innovative solutions, their benefits, but also the risks and dependencies, especially from non-European hyperscalers. The presentation will highlight possible market pathways for the future developments of Cloud-Edge-IoT infrastructure that build a backbone for Edge AI applications.
Speaker: Elisabeth Steimetz, EPoSS, VDI/VDE-IT GmbH, Germany; and Inessa Seifert, VDI/VDE-IT GmbH, Germany.
Elisabeth Steimetz studied physics at RWTH Aachen, Germany with focus on semiconductor technologies and received her Ph.D. in physics from Technical University Berlin in the field of in-situ metrology of III-V materials. For more than 20 years she was active in the field of compound semiconductors.
As a co-founder of LayTec GmbH (1999), she successfully built up the company and held the position of Director Strategic Marketing & Sales for 13 years. She is familiar with the semiconductor industries and the needs of SMEs.
In 2013 Elisabeth joined VDI/VDE Innovation and Technology GmbH as project officer for national and European funded projects in the field of Electronic Components and Systems. From 2014 until end of 2018 she was a German delegate in the public authority board of the EUREKA Clusters CATRENE and later PENTA, were she acted as national contact point as well.
Since 2019 she serves as Director for EPoSS – the European Platform on Smart Systems Integration – operated by VDI/VDE IT. In this position she was Chairperson for the ECS-SRA 2020 and ECS-SRIA 2021. She is lead delegate for EPOSS in the Private members Board and Governing Board of the Chips JU – the Pillar 1 activities of the European Chips Act.
16:30-16:50
Presentation: E01-4: PRESENTATION
Presentation: F01-4: EDGE-ORIENTED DISTRIBUTED ARCHITECTURE BASED ON ROAD-SIDE-PERCEPTION-UNITS (RSPUS) TO SUPPORT A FEDERATED LEARNING APPROACH FOR URBAN AIR QUALITY PREDICTIONS.
George Bravos, ITML, Greece.
A comprehensive edge-oriented distributed architecture is proposed for the acquisition of air quality measurements at the micro-edge level and its distribution to a network of RSPU devices at the deep edge level over a wireless Lora mesh. Each RSPU device aggregates data from multiple sensor nodes and performs local ML training and inference for Air Quality Index predictions. All deployed RSPUs are connected at the meta-edge level via a dedicated Federated Learning approach for combining locally trained ML models and distributing them to the RSPUs for achieving a complete end-to-end distributed edge AI system.
Speaker: George Bravos, ITML, Greece.
George Bravos is the co-founder of ITML and the Director of Research and Development. He received his Electrical Engineering and Computer Science diploma from the National Technical University of Athens (NTUA) at 2002 and his PhD from the University of Piraeus at 2008. He has a solid background as a research scientist and an innovator. From 2002 to 2008 he worked as a research engineer at the University of Piraeus in several international research projects. From 2009 to 2014 he worked as a scientific assistant – professor at the Technical Educational Institution of Chalkida, Greece. Since 2016 he has successfully managed more than a hundred (100) National and EU funded research and innovation projects. His research interests mainly focus on physical layer techniques and MAC protocols for energy efficient wireless sensor networks. He is a member of IEEE and a member of the Technical Chamber of Greece.
17:00
End
18:00
Conference Cocktail – Grotta Marcello, Cagliari
Note: The links to the talks are available under the drop-down menu of each presentation description.
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
8:30-17:00
Tuesday 22 October 2024 – Day 2
9:00-10:30
Plenary Session II (Room Castello):European Edge AI Ecosystem
9:00-09:30
AI AND SAFETY AS KEY ELEMENTS FOR A SOFTWARE DEFINED VEHICLE
Patrick Pype, CTO, NXP Semiconductors, Belgium.
The European automotive industry is facing several challenges which include electrification, autonomous driving, connectivity with the cloud and service oriented embedded solutions. A key element in this transition process is the need to develop new software platforms that can support the increasing complexity of modern cars. Vehicles will be permanently embedded in ubiquitous cloud structures; they will be integrated into the infrastructure of smart cities and intelligent highways and will be represented in cyberspace as digital (real-time) twins. Within Europe, the European automotive industry (OEMs, tiers and tool suppliers) work in close collaboration with the European semiconductor and embedded software industry to develop non-differentiating building blocks of future Software Defined Platform Architectures. Important elements in this development are the continuous attention of safety and the integration of AI to create the necessary agility required to catch up and be successful in the competition to success in the software-defined vehicles.
Speaker: Patrick Pype, CTO, NXP Semiconductors, Belgium.
Patrick Pype is Director of Strategic Partnerships at NXP Semiconductors and has held this position since 2012. He began his career at IMEC in Belgium in 1986. In 1996 he started the start-up company CoWare on hardware/software co-design which was later acquired by Synopsys. Thereafter he worked at Philips in different positions.
9:30-10:00
OVERVIEW OF THE INDUSTRIAL ASSOCIATIONS ACTIVITIES
Elisabeth Steimetz, EPoSS Industry Association, VDI/VDE-IT GmbH, Germany; and Paolo Azzoni, Inside Industry Association, the Netherlands
The presentation will give an overview of the activities of the AENEAS, EPoSS and Inside. The industry associations AENEAS, EPoSS, and INSIDE, constituting the private members of the Chips Joint Undertaking, would like to put forward some considerations and recommendations, primarily from an industrial perspective.
Speaker: Elisabeth Steimetz, EPoSS Industry Association, VDI/VDE-IT GmbH, Germany; and Paolo Azzoni, Inside Industry Association, the Netherlands.
Elisabeth Steimetz studied physics at RWTH Aachen, Germany with focus on semiconductor technologies and received her Ph.D. in physics from Technical University Berlin in the field of in-situ metrology of III-V materials. For more than 20 years she was active in the field of compound semiconductors. As a co-founder of LayTec GmbH (1999), she successfully built up the company and held the position of Director Strategic Marketing & Sales for 13 years. She is familiar with the semiconductor industries and the needs of SMEs. In 2013 Elisabeth joined VDI/VDE Innovation and Technology GmbH as project officer for national and European funded projects in the field of Electronic Components and Systems. From 2014 until end of 2018 she was a German delegate in the public authority board of the EUREKA Clusters CATRENE and later PENTA, were she acted as national contact point as well. Since 2019 she serves as Director for EPoSS – the European Platform on Smart Systems Integration – operated by VDI/VDE IT. In this position she was Chairperson for the ECS-SRA 2020 and ECS-SRIA 2021. She is lead delegate for EPOSS in the Private members Board and Governing Board of the Chips JU – the Pillar 1 activities of the European Chips Act.
Paolo Azzoni is the Secretary General of INSIDE Industry Association (formerly Artemis-IA), the industry association that serves as the European Technology Platform for research, design and innovation on Intelligent Digital Systems and their technology ecosystems. INSIDE is one of the three private members of the Chips-JU, the tripartite partnership between the European Commission, the Participant States and private entities, mobilizing more than 11 billion euro to safeguard, consolidate, and strengthen the Electronic Components and Systems value chain in Europe. In this context, he is the lead delegate in the Chips Ju Governing Board and the co-chairman of the ECS Strategic Research and Innovation Agenda, a funding-agnostic document describing the major challenges and priorities in the Electronic Components and Systems domain for the next 10 years. He is also the Head of European Technology Programmes at EUROTECH Group, planning and directing industrial research projects, in the areas of cyber-physical systems, intelligent systems, machine-to-machine technologies, edge computing, internet of things and digitalization solutions. Before joining EUROTECH, he was involved in academic lecturing and research in the areas of hardware formal verification, hardware/software co-design and co-simulation, advanced hardware architectures and operating systems. He holds a Master Degree in Computer Science and a second Master Degree in Intelligent Systems.
10:00-10:30
Project Pitch Presentations – AI4CSM, A-IQ Ready, LoLiPoP-IoT, ARCHIMEDES, CLEVER. (Room Castello)
10:30-11:00
Morning Coffee/Tea Break (Room Terrazza)
11:00-17:00
Parallel Sessions
11:00-12:30
Session A02 (Room Stampace): TRANSFORMERS AT THE EDGE.
Session Co-chairs:
Patrick Pype, NXP Semiconductors, Belgium; Konstantinos Georgopoulos, Technical University of Crete and National Technical University of Athens, Greece.
The session examines the integration of transformer models in edge computing environments, highlighting their potential to revolutionise real-time data processing and decision-making. Speakers will explore the unique challenges of deploying large-scale transformer architectures on resource-constrained devices and strategies for model compression, quantisation, and optimisation. They will present case studies showcasing successful applications in natural language processing. This session aims to inspire attendees by demonstrating how transformers can enhance edge AI capabilities, making intelligent solutions more accessible and efficient.
Session B02 (Room Castello): NEUROMORPHIC HW ARCHITECTURES.
Session Co-chairs:
Björn Debaillie, IMEC, Belgium; Fabian Chersi, CEA-LIST, France.
The session explores innovative designs that mimic the structure and functionalities of the human brain, aiming to achieve energy-efficient and highly parallel processing for AI applications. Speakers will present the principles behind neuromorphic computing, examining various architectures and their suitability for tasks such as sensory data processing and cognitive computing. They will discuss the advantages of neuromorphic systems, including low power consumption and real-time learning capabilities, and present case studies demonstrating their effectiveness in robotics, healthcare, and IoT. This session seeks to inspire new approaches in AI hardware design and implementation.
11:00-11:20
Presentation: A02-1: EFFICIENT QUANTIZATION OF VISION TRANSFORMERS FOR MOBILE DEVICES INFERENCE: CHALLENGES AND POST-TRAINING STRATEGIES.
Martyna Poreba, CEA-List, Université Paris-Saclay, France.
The emergence of transformer networks raises huge expectations, promising unequalled performance and generalization properties, making them suitable for a wide variety of application domains. However, these models come with their own implementation issues, particularly in a world in search of frugality. The ability to run efficiently these models inference on mobile devices is therefore essential to deploying them on a large scale, while controlling the energy cost. The key challenge is to balance application efficiency with the compression rate of the handled data. In the case of vision transformers (ViT), the heterogeneity and successive processing operations require a variety of fine-tuned approaches. In this article, we propose a state-of-the-art review of the various quantization methods applied to these networks to optimize hardware inference. We particularly focus on the impact of identified methods on classification task accuracy and the need for fully integer processing pipeline. In conclusion, we suggest a co-design Post-Training Quantization evaluation scheme for ViT. The proposed approach aims to identify the most suitable fully integer quantization strategy based on the exploration of heterogeneous combinations of relevant existing methods.
Speaker: Martyna Poreba, CEA-List, Université Paris-Saclay, France.
Martyna Poreba is a research engineer at CEA-List, specializing in multimodal perception for embedded systems. She earned her PhD in Computer Science from MINES ParisTech in 2014. Martyna worked then at the French Ministry of Ecology, developing a cross-domain, multi-source image retrieval engine to valorize France’s environmental heritage. In recent years, she has shifted her focus toward deep learning, with particular expertise in visual-inertial navigation systems (VINS) and optimizing Vision Transformers (ViT) for dense prediction tasks like object detection and semantic segmentation. Her research is driven by a commitment to advancing AI technologies for practical, real-world applications.
Presentation: B02-1: NEUROMORPHIC ARCHITECTURE FOR IOT EDGE WATER MANAGEMENT.
Mugdim Bublin and Heimo Hirner, University of Applied Science FH Campus Wien, Austria.
The exponential growth of IoT networks necessitates a paradigm shift towards architectures that offer high flexibility and learning capabilities while maintaining low energy consumption, minimal communication overhead, and low latency. Traditional IoT systems, particularly when integrated with machine learning approaches, often suffer from high communication overhead and significant energy consumption. This work addresses these challenges by proposing a neuromorphic architecture inspired by biological systems. To illustrate the practical application of our proposed architecture, we present a case study focusing on water management in the Carinthian community of Neuhaus. Preliminary results regarding water consumption prediction and anomaly detection in this community are presented. We also introduce a novel neuromorphic IoT architecture that integrates biological principles into the design of IoT systems. This architecture is specifically tailored for edge computing scenarios, where low power and high efficiency are crucial. Our approach leverages the inherent advantages of neuromorphic computing, such as asynchronous processing and event-driven communication, to create an IoT framework that is both energy-efficient and responsive. This case study demonstrates how the neuromorphic IoT architecture can be deployed in a real-world scenario, highlighting its benefits in terms of energy savings, reduced communication overhead, and improved system responsiveness.
Speaker: Mugdim Bublin and Heimo Hirner, University of Applied Science FH Campus Wien, Austria.
Mugdim Bublin studied Electrical Engineering at the Technical University of Vienna, earning his DI in 1997 and PhD in 2004. He spent almost 20 years at Siemens, focusing on Radio Resource Management Algorithms for 3G and 4G mobile networks, patent management, and machine learning. He also contributed to automotive software safety at Bosch. Since 2019, he has been affiliated with the University of Applied Sciences, FH Campus Wien. In 2021, he was appointed the City of Vienna Endowed Professor of Artificial Intelligence. His extensive research has resulted in numerous publications and patents, particularly in the application of AI in mobile networks, IoT, sensor technology and industry.
Heimo Hirner studied Technical Physics at the Technical University of Vienna. He worked in industry until 2017 in the fields of optical research at Philips and Software Integration in telecommunications at Alcatel-Lucent and Nokia. He has been working at the University of Applied Sciences FH Campus Wien since 2017 in the Department of Computer Science. Since 2022, he has headed the Research Center AI, Software and Safety at FH Campus Wien.
11:20-11:40
Presentation: A02-2: OPTIMIZING VISION TRANSFORMERS FOR EDGE DEPLOYMENT: HYBRID TOKEN REDUCTION FOR EFFICIENT SEMANTIC SEGMENTATION.
Martyna Poreba, French Alternative Energies and Atomic Energy Commission (CEA), France.
Vision Transformers (ViTs) excel in accuracy for multiple vision-related tasks; however, substantial computational and memory demands limit their deployment on resource-constrained edge devices. ViTs process images by dividing them into uniform patches, each treated as a token. Since not all regions are equally important —detailed areas may require more tokens, while broader regions need fewer—optimizing token processing is essential for improving efficiency. To enhance ViT computational performance, we implement a hybrid token reduction approach, integrating token merging and pruning strategies within MMSegmentation, a widely-used open-source semantic segmentation toolbox. Our method combines the strengths of CTS, which merges semantically similar and adjacent tokens using a CNN-based policy network, and DToP, which halts the processing of tokens that can be predicted with sufficient accuracy in the early layers of the network. The experimental results show up to a 2× reduction in computational complexity with only an approximate 1% mIoU drop on the NVIDIA Jetson AGX Orin 64GB. While DToP is integrated into the PyTorch-based MMSegmentation toolbox, exporting the pruned PyTorch model to TensorRT remains a challenging task that requires significant additional effort. We highlight these issues and outline the further work needed to ensure full compatibility with ONNX export standards.
Speaker: Martyna Poreba, French Alternative Energies and Atomic Energy Commission (CEA), France.
Martyna Poreba is a research engineer at CEA-List, specializing in multimodal perception for embedded systems. She earned her PhD in Computer Science from MINES ParisTech in 2014. Martyna worked then at the French Ministry of Ecology, developing a cross-domain, multi-source image retrieval engine to valorize France’s environmental heritage. In recent years, she has shifted her focus toward deep learning, with particular expertise in visual-inertial navigation systems (VINS) and optimizing Vision Transformers (ViT) for dense prediction tasks like object detection and semantic segmentation. Her research is driven by a commitment to advancing AI technologies for practical, real-world applications.
Presentation: B02-2: MEMRISTIVE ASSOCIATE-LEARNING ARCHITECTURE FOR HIGH-ACCURACY ULTRALOW POWER EDGE LEARNING.
Sayani Majumdar, Tampere University, Finland.
Real-time learning and decision making in hardware is essential for energy-efficient edge devices. Associative learning is a fundamental learning process by which biological systems form connections between correlated events, enabling organisms to make decisions based on prior experiences. Creating an electronic analog of associative learning circuit with emerging synaptic components like memristors can provide vital information on how emerging devices would behave in real-time operating neuromorphic circuits. The current study focuses on the associative learning behavior observed in dogs, known as Pavlovian learning, where the sound of a bell becomes associated with food, triggering salivation based on past memories. However, existing analogue circuit models that aim to replicate this phenomenon have encountered difficulties, particularly in distinguishing between different stimuli and erroneously exhibiting learning without actually achieving it. Consequently, these circuits demonstrate associative learning only under specific signal sequences, failing to exhibit genuine associative learning when stimuli change orders. To address these challenges, a novel circuit configuration is proposed, utilizing mixed-signal and non-volatile memory elements. This innovative design overcome limitations related to signal sequencing, enabling the circuit to demonstrate true associative learning regardless of the order in which stimuli are presented, while also emulating additional brain functions such as forgetting.
Speaker: Sayani Majumdar, Tampere University, Finland.
Sayani Majumdar is an Associate Professor of Electrical Engineering at the Faculty of Information Technologies and Communication Sciences at Tampere University. She is a Steering committee member of the System-on-Chip (SoC) hub at Tampere University. She is also a visiting professor at the Miin Wu School of Computing at National Cheng Kung University, Taiwan. Her career in semiconductor device research spans more than two decades where she worked in both academia and in close collaboration with the industry. She worked on several emerging technologies in world-leading organizations including MIT, USA and Max-Planck Institute, Germany. Her team’s current research focuses on the development of neuromorphic computing chips using low-thermal budget materials and processes for their back-end-of-line integration with CMOS circuits, neuromorphic architectures and their application in edge devices.
11:40-12:00
Presentation: A02-3: OPTIMIZING COMPILATION STRATEGIES WITH DeepCodeGraph: A GRAPH NEURAL NETWORK APPROACH TO SOURCE CODE ANALYSIS.
Francesco Barchi, University of Bologna, Italy.
Analysing source code using deep learning aids compile-time decisions affecting performance in embedded devices. We propose DeepCodeGraph (DCG), a general graph-based language model, which learns patterns to identify better compilation strategies, optimal hardware configurations and software transformations. DCG includes i) A large-scale dataset containing over 100k graphs. ii) A graph neural network to implement a graph-based language model. iii) A self-supervised pre-training framework leveraging Masked Graph Autoencoders. The performance of DCG is evaluated on two downstream tasks: heterogeneous device mapping and thread block size prediction. DCG outperforms previous graph-based state-of-the-art improving previous results by 3%.
Speaker: Francesco Barchi, University of Bologna, Italy.
Francesco Barchi is a researcher at the University of Bologna, he conducts his research in the field of embedded systems. His research interests focus, in particular, on using machine learning to optimise compilation phases and on optimisation problems for cyber-physical systems (CPS) composed of heterogeneous architectures.
He obtained his PhD in Computer Engineering at the Polytechnic University of Turin. During his PhD, he worked on mapping Spiking neural networks on the SpiNNaker neuromorphic platform by developing an optimised communication middleware for the same architecture.
He currently focuses his research on vertical applications of embedded architectures involving infrastructure monitoring and the development of security middleware for RISC-V architecture-based SoCs for cyber-physical systems.
Presentation: B02-3: ENABLING NEAR-SENSOR AI WITH LOW-POWER PROCESSING: AN OPEN-HARDWARE EXPERIENCE.
Luca Martis, Università degli Studi di Cagliari, Italy.
Moving data processing closer to its source, known as edge computing, offers significant advantages including reduced latency, bandwidth savings, and enhanced reliability. However, implementing complex algorithms on edge devices is challenging due to power and computational limitations. Moreover, for most actors on the market, accessing chip fabrication has significant cost barriers. We present our experience gained during the implementation of an accelerator for Spiking Neural Networks designed to be integrated into a System-on-Chip (SoC). The accelerator was developed using the open-source Process Design Kit (PDK) SKY130A and open-source Electronic Design Automation (EDA) tools. Preliminary results highlight the potential of this open instruments in EdgeAI applications, in terms of performance, power and chip-making democratization, as well as the current limitations.
Speaker: Luca Martis, Università degli Studi di Cagliari, Italy.
Luca Martis is a PhD student at the University of Cagliari, where he earned his MSc in 2023 and BSc in 2021. His research interests include the development of custom hardware systems for executing artificial intelligence algorithms, with a particular focus on Spiking Neural Networks.
Presentation Link
12:00-12:20
Presentation: A02-4: EXPLORING THE BENEFITS OF VISION FOUNDATION MODELS FOR UNSUPERVISED DOMAIN ADAPTATION.
Tommie Kerssies, Eindhoven University of Technology, The Netherlands.
Robust generalization across diverse data domains is a critical challenge in computer vision, especially in safety-critical applications. Our study explores how combining Vision Foundation Models with Unsupervised Domain Adaptation techniques can enhance semantic segmentation performance. Our findings show that this integration not only improves adaptability to unseen environments but also reduces computational complexity by eliminating redundant components. These results underscore the significant benefits of combining Vision Foundation Models with Unsupervised Domain Adaptation, setting new standards and baselines for unsupervised domain adaptation in semantic segmentation.
Speaker: Tommie Kerssies, Eindhoven University of Technology, The Netherlands.
Tommie Kerssies is a PhD candidate in the Mobile Perception Systems Lab, specializing in Vision Foundation Models and their application in image segmentation and domain generalization.
Presentation: B02-4: PRIVACY-PRESERVING DISTRIBUTED LEARNING AT THE EDGE THROUGH APPROXIMATE COMPUTING.
Ihsen Alouani, CNRS, Université Polytechnique Hauts-de-France.
In this presentation, we introduce a novel approach to privacy-preserving distributed learning in the Edge, leveraging the potential of approximate computing within a federated learning framework. Our system deploys edge clients that operate on approximate computing, enhancing both performance and privacy preservation with a by-product energy and resources gain. By using approximate in computing, we demonstrate an improved capacity for edge devices to protect sensitive data during the training process.
We evaluate the resilience of our system under various attack scenarios, particularly focusing on the deep leakage from gradient attack, a critical threat in federated learning environments. Our experimental results show that edge nodes equipped with approximate computing exhibit significantly higher resilience to privacy breaches compared to those relying on conventional, exact computing hardware. This work highlights the potential of approximate computing not only to improve efficiency but also to fortify the privacy of edge-based federated learning systems in adversarial settings.
Speaker: Ihsen Alouani, CNRS, Université Polytechnique Hauts-de-France.
Ihsen Alouani received his master’s degree from the National Engineering School of Sousse, Tunisia, and completed his Ph.D. at Université Polytechnique Hauts-de-France in 2016. Following his Ph.D., he undertook a postdoctoral position at the Technical University of Munich, Germany. Since 2018, he has been an Associate Professor at the IEMN Laboratory (CNRS) at Université Polytechnique Hauts-de-France. In 2022, he joined Queen’s University Belfast, affiliated with the Centre for Secure Information Technologies (CSIT).
Alouani is the Principal Investigator of several projects from different funding agencies (EPSRC, EU projects, etc). He is serving on the program committees of several international conferences, including the Design Automation Conference (DAC), Asia and South Pacific Design Automation Conference (ASP-DAC), etc. His research interests include machine learning security and privacy, approximate computing, and the security of embedded systems.
12:30-13:30
Networking Lunch (Room Terrazza)
13:30-15:30
Session C02 (Room Stampace): EDGE AI HARDWARE-SOFTWARE CO-DESIGN METHODS AND TOOLS.
Session Co-chairs:
Marcello Coppola, STMicroelectronics, France; Angelo Genovese, Università degli Studi di Milano, Italy.
The session focuses on the critical hardware and software integration in developing efficient edge AI solutions. Participants will explore innovative co-design methodologies that optimise hardware components and software algorithms, enhancing performance while minimising resource consumption. Speakers will discuss various design tools and frameworks that facilitate seamless communication between hardware and software, ensuring compatibility and efficiency. This session will equip attendees with practical strategies to accelerate their edge AI development processes.
Session D02 (Room Castello): RECONFIGURABLE EDGE AI.
Session Co-chairs:
Heimo Hirner, University of Applied Science FH Campus Wien, Austria; Alberto Faro, Deepsensing SRL, Italy.
The session delves into the adaptability and flexibility of edge computing systems designed to meet dynamic application requirements in real-time. Speakers will explore reconfigurable architectures, including Field-Programmable Gate Arrays (FPGAs) and adaptable software frameworks, that enable efficient deployment and scaling of AI models at the edge. They will discuss the advantages of reconfigurability in enhancing performance, optimising resource utilisation, and supporting diverse tasks. By presenting case studies and best practices, this session aims to inspire innovative approaches and solutions for implementing reconfigurable edge AI systems.
13:30-13:50
Presentation: C02-1: MANDALA: OPTIMISED ARTIFICIAL INTELLIGENCE HPDP INTERFACE USING KLEPSYDRA AI STREAMING.
Pablo Ghiglino, Klepsydra Technologies, Switzerland.
The reliance on radiation-hardened hardware, crucial for high-dependability fields like space, automotive, and medical applications, limits the selection of components for AI-intensive tasks, particularly in real-time classifications. To address this, we propose using the High-Performance Data Processor (HPDP) as a co-processor, combined with an AI framework optimized for data processing. The HPDP’s dynamic reconfiguration and dataflow architecture, paired with Klepsydra’s AI-runtime inference framework, enable efficient, low-latency processing of continuous AI-driven applications. This integration, along with the Streaming Distribution Optimizer (SDO), creates a robust dataflow system. Our approach eliminates the need for HPDP-specific coding by porting the AI framework and SDO to the HPDP. This paper presents our implementation and performance evaluation, highlighting the HPDP as a reliable, reprogrammable co-processor suitable for AI deployment in challenging environments like space exploration. Our proposed solution offers a significant advancement in deploying AI on radiation-hardened platforms, by using the HPDP as a dependable co-processor, which is not only efficient but also reprogrammable, making it highly suitable for any application requiring dependable execution in any environment such as planetary landers and deep space robotics. Therefore, this fully European solution also strengthens Europe’s position in AI-driven space exploration.
Speaker: Pablo Ghiglino, Klepsydra Technologies, Switzerland.
Pablo Ghiglino has over 10-year experience in the finance industry (e.g. at Bank of America, Credit-Suisse and HSBC in London, Julius Baer in Zurich) as a software developer. This experience exposed him to this high-intensity and high-speed data-processing environment in which data loss results in high economic impact. Parallel to his work in finance, Pablo successfully studied a PhD at the Surrey Space Center in Aerospace Engineering in the field of “High Precision Planetary Landing”. After his studies, he got involved in several navigation projects, where he seized the opportunity and created Klepsydra to respond to the needs for cutting embedded computing by implementing successful software techniques from the highly competitive financial sector in embedded software frameworks for its use in other industries.
Presentation: D02-1: SWARM NETWORKING FOR SMART FACTORY: THE SMARTEDGE SOLUTION.
Mohamed Mtiri, CNIT, Italy.
An innovative swarm networking solution is designed and implemented in the context of a decentralized smart factory scenarios including mobile robots. The solution relies on P4-based network programmability to provide secure swarm formation and swarm management.
Speaker: Mohamed Mtiri, CNIT, Italy.
Mohamed Mtiri is Researcher at CNIT, Pisa, Italy. His main interests include programmable networking, cyber security, and AI technologies for communication.
13:50-14:10
Presentation: C02-2: EDGEAI: LEVERAGING SNNS, TRANSFORMERS, AND AUTOENCODERS FOR NEAR-SENSOR INTELLIGENCE.
Paola Busia, Università degli Studi di Cagliari, Italy.
This talk will explore the cutting-edge advancements in EdgeAI for biosignal analysis, focusing on innovative algorithms such as Spiking Neural Networks (SNNs), Transformers, and Autoencoders. As the demand for real-time, near-sensor processing of biosignals like EEG, ECG, and EMG continues to grow, traditional cloud-based methods face challenges related to latency, energy efficiency, and data privacy. We will discuss how SNNs, with their neuromorphic efficiency, Transformers, with their powerful sequence modelling, and Autoencoders, with their capacity for anomaly spotting, are being adapted for edge deployment. The talk will highlight real-world applications inside and outside the wearable healthcare domain, demonstrating how these technologies enable fast, accurate, and energy-efficient processing at the edge.
Speaker: Paola Busia, Università degli Studi di Cagliari, Italy.
Paola Busia received the Ph.D. degree in electronics and computer engineering from the Università degli studi di Cagliari, Cagliari, Italy, in 2023. She is currently an Assistant Professor with the Università degli studi di Cagliari. Her main research interests include power efficient processing architectures for embedded applications and for at-the-edge AI.
Presentation: D02-2: AI-BASED DYNAMICALLY CONTROL AND OPTIMIZE LIGHTING CONDITIONS BASED ON AN AMBIENT INPUT.
Tore Karlsen, ProLux AS, Norway.
The presentation addresses the concept and design for a system using edge AI with lighting systems to achieve energy efficiency, improve user experience, and respond to environmental factors.
Speaker: Tore Karlsen, ProLux AS, Norway.
Tore Karlsen (CTO) has his degree in electronics and automation and gained competence and experience in project management, product development, system development and project planning through a 25+ year long career in different engineering roles. The field of competence stretches over technical architecture design and project planning with responsibilities of both technical and economical perspectives.
14:10-14:30
Presentation: C02-3: AUTOMATIC HARDWARE-AWARE DESIGN, OPTIMIZATION, AND DEPLOYMENT OF DEEP LEARNING MODELS ON HETEROGENEOUS EDGE DEVICES.
Matteo Risso. Politecnico di Torino, Italy.
This abstract presents an innovative flow for optimizing and deploying deep neural networks (DNNs) on constrained edge devices, leveraging two complementary in-house developed open-source tools: PLiNIO and MATCH. PLiNIO is a PyTorch-based package that enables efficient DNN inference optimization with minimal code changes. It employs gradient-based techniques like Differentiable Neural Architecture Search (DNAS) and Mixed-Precision Search (MPS) to optimize model architectures while considering hardware-specific constraints. PLiNIO’s cost models account for various metrics, from basic parameter counts to advanced hardware-specific features like spatial parallelism and dataflow patterns. MATCH complements PLiNIO by translating optimized models into deployable C code. As a Model-Aware TVM-based Compiler for Heterogeneous hardware, MATCH exploits Apache TVM’s BYOC framework to efficiently map DNNs onto diverse edge System-on-Chip (SoC) platforms. It partitions networks to match hardware-specific patterns, manages accelerator-specific compilation, and leverages TVM for unsupported layers. MATCH’s temporal mapping engine, powered by ZigZag, optimizes scheduling and memory allocation for each layer. Together, PLiNIO and MATCH form a comprehensive, free and open-source solution for hardware-aware DNN optimization and deployment. This flow enables researchers and developers to automatically design, refine, and implement efficient deep learning models tailored for resource-constrained edge devices, balancing accuracy with hardware-specific performance and energy efficiency constraints.
Speaker: Matteo Risso. Politecnico di Torino, Italy.
Matteo Risso received his B.Sc degree in Physical Engineering and M.Sc degree in Electronic Engineering at the Politecnico di Torino, Italy, in 2018 and 2020. He is currently working toward his Ph.D. degree at Politecnico di Torino, Italy. His research interests include Embedded Machine and Deep Learning, Design Automation, and Energy-Efficient Embedded Systems.
Presentation: D02-3: AN ADAPTIVE FORECASTING PLATFORM ENABLING DIGITAL TWINNING
Luca Valcarenghi Scuola Superiore Sant’Anna, Italy.
The proposed method assures data confidentiality through permutation and Principal Component reduction of the monitored data. Failure detection leverages clustering algorithms to differentiate optical network states. Results show that the proposed method performs well for small data size, that is suitable for real-time failure detection.
Speaker: Luca Valcarenghi Scuola Superiore Sant’Anna, Italy.
Luca Valcarenghi is an Associate Professor at the Scuola Superiore Sant’Anna of Pisa, Italy, since 2014. He received the Laurea in Electrical Engineering in 1997 from Politecnico di Torino and the M.S.E.E. and Ph.D. in Electrical Engineering Major Telecommunications from UTD in 1999 and 2001, respectively. He published more than three hundred papers in International Journals and Conference Proceedings. Valcarenghi received a Fulbright Research Scholar Fellowship in 2009 and a JSPS “Invitation Fellowship Program for Research in Japan (Long Term)” in 2013. His main research interests are optical networks design, analysis, and optimization; communication networks reliability; energy efficiency in communications networks; optical access networks; zero touch network and service management; 5G technologies and beyond.
14:30-14:50
Presentation: C02-4 – PREVAIL – OPEN ACCESS PILOT LINE FOR EDGE AI CHIP PROTOTYPING.
Sergio Nicoletti, CEA, France .
The PREVAIL project will establish and start operating the core of a networked, multi-hub platform providing prototype chip fabrication capability, in advanced to EU stakeholders for Artificial Intelligence (AI) applications. In this project, four major European RTOs jointly work to create a multihub Test and Experimentation Facility for edge AI Hardware (TEF HW AI) build on their advanced 300 mm fabrication, design, and test facilities. Within PREVAIL, companies and research institutions will be able to develop, test and experiment with AI product prototypes based on innovative technologies developed by the four RTOs. The technologies made available are the most promising to address the specifications (power consumption, computing performance, data handling and security) required by AI chips in general and edge AI in particular.
Speaker: Sergio Nicoletti, CEA-Leti, France.
Sergio Nicoletti Ph.D joined the CEA in 2006 to develop optical gas sensors. Nicoletti has since served as chief researcher and project leader in several industrial and collaborative projects. Currently, he coordinates the TEF-Prevail Project. He earned a doctorate in physics from Université Joseph Fourier in Grenoble, France.
Presentation: D02-4: CLOUD-EDGE CONTINUUM OF AI – ENABLED BODY SENSORS FOR MULTISPORT TRAINING, HEALTHY LIVING, AND REMOTE PATIENT MONITORING IN EUROPE.
Juan A Montiel-Nelson, Univ. Las Palmas, Spain.
As society ages, increasing strain is placed on healthcare systems and individuals, particularly those facing age-related health issues. A key biological marker of aging is the shortening of telomeres, protective caps on the ends of chromosomes. Telomere shortening accelerates cellular senescence, apoptosis, and oncogenic transformation, leading to a decline in health and lifespan. The shorter the telomere, the higher the risk of disease and lower the chances of survival. However, lifestyle interventions such as regular exercise, a balanced diet, and effective stress management can slow telomere shortening and promote longevity. Furthermore, advancements in technologies that enable regular physical activity, remote-assisted living, and post-hospitalization recovery at home are essential for supporting an aging population.
Speaker: Juan A Montiel-Nelson, Univ. Las Palmas, Spain.
Juan A. Montiel–Nelson received the M.S. degree in electrical engineering (1991) and the Ph.D. degree (1994) from the University of Las Palmas de Gran Canaria, Spain. He has authored or coauthored more than 190 papers in edited books, international journals, and conference proceedings.
He is recipient of the Myril B. Reed Best Paper Award, 2008 IEEE 51st MWSCAS International Midwest Symposium on Circuits and Systems (MWSCAS), and he is member of the MWSCAS Steering Committee, member of the IEEE Sensors, and reviewer in IEEE journals and conferences as: IEEE Transactions on Circuits and Systems I and II, IEEE Transactions on Very Large Scale Integration (VLSI).
He has broad experience from several universities and companies worldwide (Edith Cowan University, Michigan Technological University, Arizona State University, Yonsei University, Ford Greenfield Labs, Vitesse Semiconductor Corporation, etc.), and is currently professor at the University of Las Palmas de Gran Canaria.
Research field contributions includes: very high–speed integrated circuit (VHSIC) design in GaAs, SiGe, InP, and very deep submicrometer CMOS technologies; performance estimation and optimization of digital integrated circuits and systems; and verification of circuits and systems. His current research area is design/integration of biological and bio-medical sensor devices and embedded intelligence.
15:00-15:30
Afternoon Coffee/Tea Break (Room Terrazza)
15:30-17:00
Session E02 (Room Stampace): EDGE AI ARCHITECTURES, FRAMEWORKS, AND MODEL OPTIMISATION.
Session Co-chairs:
Iakovos Mavroidis, Technical University of Crete, Greece; Martyna Poreba, CEA-List, France.
The session examines the essential components and strategies for efficiently deploying AI at the edge. Speakers will explore edge architectures tailored for specific applications and the frameworks that facilitate model deployment and integration. Optimisation techniques, including quantisation, pruning, and hardware-aware training, will be emphasised, which enhance model performance while reducing latency and resource consumption. Through case studies and practical examples, this session aims to provide attendees with actionable insights and tools to effectively advance their edge AI initiatives.
Session F02 (Room Castello): EDGE AI DEPENDABLE SYSTEMS.
Session Co-chairs:
Gleb Radchenko, Silicon Austria Labs; Juan A Montiel-Nelson, Univ. Las Palmas, Spain.
The session focuses on the critical importance of reliability, safety, and security in deploying AI solutions at the edge. Participants will explore methodologies for designing dependable edge AI systems that can withstand environmental challenges, data anomalies, and cyber threats. Speakers will discuss best practices for robustness, fault tolerance, and real-time monitoring, ensuring systems operate effectively under varying conditions. Through case studies and expert insights, this session aims to equip attendees with strategies to enhance the dependability of their edge AI applications, fostering trust in technology across diverse industries.
15:30-15:50
Presentation: E02-1: A NOVEL HIERARCHICAL APPROACH TO PERFORM ON-DEVICE ENERGY EFFICIENT FAULT CLASSIFICATION.
Devesh Vashishth, Ferdinand Steinbeis Institute, Germany; Julio Wissing, Fraunhofer Institute for integrated circuits IIS, Germany.
Using convolutional neural networks (CNN) to discover sensor data patterns helps predict upcoming failures in industrial machines. Traditionally, this pipeline is deployed on the cloud as deep neural networks have high computational requirements. Alternatively, an on-device deployment to make decisions locally for this pipeline could lower the energy requirements by not sending the sensor data back and forth to the cloud. However, this is challenging from an edge deployment perspective due to the limited computational resources and energy budget. To approach this issue, we propose a hierarchical architecture that leverages multiple smaller networks dividing the bigger problem into smaller sub-problems in a divide-and-conquer approach. With this architecture we achieve a 9x reduction in energy consumption going from 0.045J per inference for a non-Hierarchical CNN classifier to 0.005J. At the same time, our approach delivers low latency with comparable accuracy to the baseline, while running completely at the edge.
Speaker: Devesh Vashishth, Ferdinand Steinbeis Institute, Germany; Julio Wissing, Fraunhofer Institute for integrated circuits IIS, Germany.
Devesh graduated with his Bachelors of technology in Electronics and communication engineering from the LNM institute of information and technology in year 2018. Later he obtained his masters degree in Applied computer science from Hochschule Heidelberg in 2022. He is active in research related to energy efficient deep learning for edge devices since last 3 years.
Presentation Link
Presentation: F02-1: AN AI-BASED HARDWARE/SOFTWARE WATER HEATER SOLUTION FOR PROCESSING THE ENERGY DATA AT THE EDGE.
Mats Jørgen Øyan, Nxtech AS, Norway.
An AI-based hardware/software solution processing the energy data at the edge for water tank applications.
Speaker: Mats Jørgen Øyan, Nxtech AS, Norway.
Mats Jørgen Øyan (Software Lead) has his Master of technology, computer science from NTNU, Norway in 2006. He is currently the software lead at Nxtech, Norway focusing on the design of software for embedded systems used in different applications spanning from manufacturing, energy, mobility and healthcare.
Presentation Link
15:50-16:10
Presentation: E02-2: EDGE AI HARDWARE ACCELERATION FOR CRITICAL SYSTEMS: FROM FPGAS AUTOMATIC HARDWARE GENERATOR TO VERSATILE CGRAS SYSTEMS.
Pietro Nannipieri, University of Pisa, Italy.
This work discusses advancements in onboard AI processing for critical systems, focusing on satellites. The work explores hardware solutions for real-time and autonomous data processing using Field Programmable Gate Arrays (FPGAs) and Coarse-Grained Reconfigurable Arrays (CGRAs). The research outlines challenges faced by critical systems, such as power constraints, radiation tolerance, and the need for reliable AI processing. The presentation details the design and implementation of a GPU-based system and the FPGA-AI framework that improves AI processing in critical systems. Future research focuses on extending neural network support and exploring collaborations for tape-out opportunities of CGRA prototypes.
Speaker: Pietro Nannipieri, University of Pisa, Italy.
Pietro Nannipieri got his PhD in Information Engineering in 2020 cum laude from the University of Pisa, where he is currently an Assistant Professor. His interests are digital and VLSI design, electronics for space applications, and cryptography. In 2019, Pietro was a visiting researcher in the TEC-EDP Section in ESTEC (ESA), where he carried out different qualification tests on the SpaceFibre technology. His research focuses on hardware IPs for satellite onboard data handling, signal processing, and hardware cryptography. Indeed, he is currently working on the European Processor Initiative (EPI) project.
Presentation Link
Presentation: F02-2: FUNCTIONAL AND NON-FUNCTIONAL REQUIREMENTS IN EDGE AI SYSTEMS.
Ovidiu Vermesan, SINTEF, Norway; Alain Pagani, DFKI, Germany.
This presentation highlights the critical elements defining edge AI solutions’ performance and effectiveness. It details the concepts of functional requirements, including data processing capabilities, real-time analytics, and system integration that empower applications across various sectors. The presentation examines non-functional requirements such as scalability, reliability, security, and energy efficiency. By understanding these requirements, stakeholders can better design edge AI systems that are effective but also robust and sustainable, ensuring they meet the demands of a rapidly changing technological landscape.
Speaker: Ovidiu Vermesan, SINTEF, Norway; Alain Pagani, DFKI, Germany.
Ovidiu Vermesan is Chef Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies and applying the developments to applications such as green mobility, energy, buildings, autonomous systems, electric connected, autonomous, and shared vehicles, and industrial manufacturing. He is currently the coordinator of the Edge AI Technologies for Optimised Performance Embedded Processing (EdgeAI) project.
Alain Pagani is Principal Researcher and deputy director of the Augmented Vision research department at the German Research Center for Artificial Intelligence (DFKI). His research interests include Artificial Intelligence, Computer Vision, Image Understanding, and Extended Reality. He is the coordinator of the Network of Excellence dAIEDGE about distributed AI at the edge, and the coordinator of the Horizon Europe project CORTEX about remote cooperation using extended reality. He is lecturer at the University of Kaiserslautern-Landau, and he published over 100 articles in conferences and journals. His research finds applications in eXtended Reality for telecooperation (project CORTEX), Artificial Intelligence and Computer Vision for Human-Robot cooperation (project FLUENTLY), Artificial Intelligence and Augmented Reality for analysis of extremely large data (project ExtremeXP).
16:10-16:30
Presentation: E02-3: EFFICIENT EDGEAI COMPUTER VISION THROUGH VARIOUS MODEL OPTIMIZATION TECHNIQUES.
Gerardo Daalderop, NXP Semiconductors B.V., The Netherlands.
In this presentation, we investigate how various optimization techniques can be utilized to optimize neural networks for relevant Computer Vision use cases for optimal deployment in the edge. We introduce and discuss the principles of state-of-the-art techniques beyond Quantization like Neural Architecture Search (NAS), NN Pruning, Weight Clustering, and present how these can be combined effectively to create a framework for joint collaborative optimization to further push performance boundaries on EdgeAI devices.
Speaker:Hiram Rayo Torres Rodriguez, and Gerardo Daalderop, NXP Semiconductors B.V., The Netherlands.
Gerardo Daalderop (PhD, MBA) is a technical director at NXP Semiconductors, having in recent years built-up an AIML Innovation Center in its Central Technology Office, focussing on AIML innovations for the Edge beyond next generation products, and across segments as mobility, industrial, home/office, IoT. Before this, he has set up and led groups, roadmaps, programs and projects in the areas of highly automated driving and safe mobility, safe and secure wireless, telematics and multi-standard communication solutions, and the connected consumer, always with the guiding principle to identify and address highly valued use-cases and user-needs with differentiating technologies, tested and demonstrated in practical trial-environments. As such he also has participated in numerous national and transnational cooperation projects.
Presentation Link
Presentation: F02-3: DEPENDABLE AI AT THE EDGE.
Eike Schultz, University of Lübeck, Germany.
Artificial Intelligence (AI) becomes a central technology in a growing number of applications. For any piece of software to be a trusted component in a system, it needs to be quantifiably dependable. This is especially relevant in safety-critical applications, on which recent EU legislature puts high regulatory requirements. Dependability in this context is composed of multiple assurances a developer needs to give regarding their system. This presentation will showcase how assurances in the three central categories (reliability, safety, and security) can be given for AI on the edge.
Speaker: Eike Schultz, University of Lübeck, Germany.
Eike Schultz is a PhD student exploring the field of Trustworthy AI. They earned their M.Sc. in Robotics and Autonomous Systems at the Universität zu Lübeck in Germany in early 2024 and have since been working at the Institute for Computer Engineering at the same university. In addition to their research, Eike Schultz is also engaging in various non-government and political organisations surrounding social equity and sustainable development.
16:30-16:50
Presentation: E02-4: EVALUATION OF YOLO MODEL FOR AGRI-FOOD REQUIREMENTS AND EDGE-AI SPECIFICATIONS.
Ana Pinzari, Institute of Engineering Univ. Grenoble Alpes, TIMA, France.
The Yolo model is now widely explored in a variety of applications. For the beverage industry, a single model can be used to both predict yield and detect wine diseases. Exploring YOLO for the specificity of one application therefore involves the right choice of hardware requirements. Optimisation methods are applied to ease the deployment of AI-based solutions and make the use of AI accelerators efficient.
Speaker: Ana Pinzari, Institute of Engineering Univ. Grenoble Alpes, TIMA, France.
Ana Pinzari received her Ph.D degree in computer science at Université de Technologie de Compiègne, Compiègne, France in 2012, in Real-Time Embedded Systems. Then, she worked in a scientific research organization ECSI (European Electronic Chips \& Systems Initiative) focusing on new design methods, tools and standards for design of complex electronic systems. In 2021, she joined Grenoble INP, TIMA, as a research engineer. Her current research focuses on exploring optimisation methods for edge-AI implementation.
Presentation Link
Presentation: F02-4: A WORKFLOW PROPOSAL FOR SAFE-AI.
Suzana Veljanovska, Zurich University of Applied Sciences (ZHAW), Switzerland.
The implication of the term safe and dependable AI is reliable execution of reliable AI models. Due to the massive computing-parallelism required by many AI-based classifiers to operate under deadline constraints, achieving reliable execution is considered a challenge – but not an insurmountable one. The state of the art in the manufacture of reliable AI models is less encouraging. Regardless of the state of the art it is anticipatable that a workflow compatible with functional safety and safety of the intended function is necessary. We approach this workflow based on industry use cases offered by the REBECCA project. In this presentation we introduce the state of the art in reliable execution, reliable AI modelling and the architectures to achieve reliable execution. We discuss why we choose CNNs as an AI architecture, how we architect the CNN into a safety/reliability framework, how the reliable execution is achieved. We then discuss a proposal to reduce the additional computational complexity introduced by reliable (redundant) execution. We then discuss the general requirements of a safe-AI workflow and present a proposal. The proposal centres around the ONNX modelling concept. We present the first trials including, not always successful, efforts with state-of-the-art tools. Relevant performance characteristics resulting from application of the workflow are presented.
Speaker: Suzana Veljanovska, Zurich University of Applied Sciences (ZHAW), Zurich, Switzerland.
Suzana Veljanovska is a research assistant at Zurich University of Applied Sciences (ZHAW) working in the field of dependable AI.
17:00
End
Note: The links to the talks are available under the drop-down menu of each presentation description.
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
8:30-17:00
Wednesday 23 October 2024 – Day 3
9:00-10:30
Plenary Session III (Room Castello): Edge AI Tools
9:00-09:30
DESIGNING ACCELERATED EDGE AI SYSTEMS WITH MODEL BASED METHODOLOGY.
Petri Solanti, Siemens EDA, Germany
10:30-11:00
Morning Coffee/Tea Break (Room Terrazza)
11:00-17:00
Parallel Sessions
11:00-12:30
Session A03 (Room Stampace): EDGE AI FEDERATED LEARNING.
Session Co-chairs:
Reiner John, AVL, Austria; Silke Braun, Infineon, Austria.
The session explores the revolutionary paradigm of decentralised machine learning that allows multiple edge devices to train AI models while preserving data privacy collaboratively. Speakers will present the principles and benefits of federated learning, including reduction in data transfer, enhanced privacy, and real-time model updates. Speakers will discuss key challenges such as communication efficiency, model convergence, and handling heterogeneous data.
Session B03 (Room Castello): TRUSTWORTHY EDGE AI.
Session Co-chairs:
Markus Tauber, Research Studio Austria; Thomas Gutt, Infineon Technologies AG, Germany.
The session addresses the imperative of ensuring reliability, ethics, and transparency in AI deployed at the edge. Speakers will explore frameworks and practices that enhance the trustworthiness of edge AI systems, focusing on data integrity, algorithmic fairness, and explainability. Speakers will discuss strategies for evaluating and mitigating risks associated with bias, adversarial attacks, and privacy breaches. The session aims to equip attendees with actionable insights and tools to foster trustworthy AI implementations, ultimately building user confidence in edge technologies across diverse sectors.
11:00-11:20
Presentation: A03-1: BAYESIAN APPROACHES TO UNCERTAINTY ESTIMATION IN EDGE AI COLLABORATIVE LEARNING.
Gleb Radchenko, Silicon Austria Labs.
The rise of edge AI significantly boosted the capabilities of IoT devices, but it has also created new issues related to knowledge sharing, particularly with respect to spatiotemporal data locality in these environments. This research investigates applicability of various approaches for deploying distributed machine learning across autonomous, AI-enabled edge devices with network capabilities. A key focus is on evaluating the reliability of learning outcomes considering the diverse spatial data characteristics handled by individual agents. By using collaborative mapping as an example, we explore the use of the Distributed Neural Network Optimization (DiNNO) algorithm, enhanced with Bayesian neural networks (BNNs) for uncertainty quantification. To simulate collaborative mapping scenarios, we employ the Webots platform to create a 3D environment, where we separate the DiNNO algorithm into distinct processes to enable asynchronous network communication in distributed learning settings and incorporate distributed uncertainty estimation using BNNs.
Speaker: Gleb Radchenko, Silicon Austria Labs.Koszescha
Gleb Radchenko received his Ph.D. in Computer Science from Moscow State University. Following his doctoral studies, he became the Dean of the Faculty of Computational Mathematics and Informatics at South Ural State University in 2013. He later served as the Director of the School of Electronic Engineering and Computer Science and headed the Department of Computer Engineering, significantly contributing to the advancement of the university’s academic programs and research initiatives. In 2021, Radchenko moved to Austria to join Silicon Austria Labs in Graz as a Staff Scientist. His work focuses on Embedded AI and Distributed Computing Systems, with particular interests in Edge AI, Federated Learning, and Cloud Computing. Throughout his career, he has made numerous contributions to software engineering, distributed systems and service-oriented architectures, aiming to develop scalable and efficient computing solutions that bridge the gap between cloud and edge technologies.
Presentation: B03-1: THE ACCOUNTABILITY STRIKES BACK: DECENTRALIZING THE KEY GENERATION IN CL-PKC WITH TRACEABLE RING SIGNATURES.
Varesh Mishra, COSIC KU Leuven, Belgium.
In the context of Federated learning, it is essential to authenticate the incoming messages from various entities. Many works have utilized Certificateless Public Key Cryptography (CL-PKC) and Signature Schemes. These recent developments focus on the existence of a single fully Trusted Authority (TA) that can issue partial private signatures truthfully without any malicious intent. This assumption is not pragmatic when multiple competing entities with some assets are involved. Additionally, dependence on a single Trusted Authority introduces a single point of failure and a center of malice. In this work, first, we propose a mechanism for partial key exchange for Certificateless Public Key Cryptography (CL-PKC) employing Traceable Ring Signatures. Furthermore, we utilize the existing Blockchain or logging infrastructure to enable our model to provide accountability and disincentivization for malicious Trusted Authorities. The cryptographic tools used in this work can be parallelized to get more efficiency out of these trusted nodes. To evaluate our protocol, we also simulate it to argue for the communication and computation costs.
Speaker: Varesh Mishra, COSIC KU Leuven, Belgium.
Varesh Mishra earned a B.Tech. in Electronics and Communications Engineering from Guru Gobind Singh Indraprastha University, New Delhi, India, in 2018 and an M.Tech. in Information Security from Atal Bihari Vajpayee-Indian Institute of Information Technology and Management, Gwalior, India, in 2022. He is currently a Ph.D. student in the COSIC group at KU Leuven, focusing on distributed system security, cryptography, and privacy-enhancing technologies.
11:20-11:40
Presentation: A03-2: FEDERATED LEARNING APPLIED ON EDGE-AI HARDWARE FOR ELECTRIC VEHICLE RANGE ESTIMATION.
André Baldermann, Ostbayerische Technische Hochschule Amberg-Weiden, Germany.
This proposal outlines the implementation of a diverse hardware setup that simulates a fleet of electric cars, optimizing through federated learning on real-world data from pre-recorded test drives to achieve maximum range and minimize energy consumption. Federated Learning is applied in a manner compliant with the EU AI Act, refining a global model for each Edge device using locally trained models. The concept demonstrates the application of Federated Learning for data privacy, adhering to EU AI Act regulations. It serves as a compelling example that companies likely have no business interest in collecting client data, thus enhancing trustworthiness. Additionally, the local model is trained ad hoc, enabling AI-driver interaction to mitigate energy-intensive driving behaviour, especially after stressful situations like traffic jams. This approach can be applied in both private and fleet vehicles to encourage safe driving practices. Energy-optimized driving addresses range anxiety and contributes to the Green Deal. This work is an outcome of the AI4CSM project.
Speaker: André Baldermann, Ostbayerische Technische Hochschule Amberg-Weiden, Germany. (TBD)
André Baldermann is an AI Researcher at the University of Applied Sciences Ostbayerische Hochschule Amberg-Weiden in Germany, where he focuses on edge AI and federated learning. Prior to his transition into AI research, André spent more than 10 years Siemens as an electrical engineer, with the last 5 years leading projects in the innovation department. During this time, he gained deep expertise in cutting-edge technologies and led cross-functional teams in the development of innovative solutions. Currently, André is developing an implementation of EU AI Act-compliant privacy-preserving federated learning in a diverse hardware demonstrator, aimed at improving range estimation for electric vehicles and enhancing user-AI interaction, as part of his work in the AI4CSM project.
Presentation Link
Presentation: B03-2: LEGAL REQUIREMENTS FOR TRUSTWORTHY EDGE AI: THE CURRENT STATE-OF-PLAY AND RECENT LEGAL DEVELOPMENTS.
Lydia Belkadi, KU Leuven, Belgium.
Achieving Trustworthy Edge AI will require developers and deployers alike to navigate several regulatory frameworks over the coming years, and to consider recent legislative activities at the European Union (‘EU’) level. New legislations have expanded the scope of risk-assessments and legal requirements on the design and development of technologies beyond personal data protection. In this context, the EU legislator has acknowledged the potential of Edge AI to complement pseudonymization and encryption techniques. The recently adopted Artificial Intelligence Act and the Ethics Guidelines on Trustworthy AI are anticipated to form essential normative keystones in the EU. A better understanding of the legal requirements applying to connected products and their interplay with the Artificial Intelligence Act will be essential for the successful uptake of AI at the Edge. This contribution examines key regulatory aspects of the Artificial Intelligence Act (‘AIA’), the Cyber Resilience Act (‘CRA’), the Data Act (‘DA’), and the General Data Protection Regulation (‘GDPR’) to propose some reflections on recent legal developments and their significance for Edge AI technologies.
Speaker: Lydia Belkadi, KU Leuven, Belgium.
To be provided.Lydia Belkadi is a doctoral researcher within the Centre for IT & IP Law at KU Leuven, Belgium. Lydia is an alumni of the H2020 MSCA-ITN TReSPAsS on privacy-preserving biometrics. She is now involved as a legal researcher within the European Network of Excellence dAIEDGE, where she researches the legal requirements for Edge AI stemming from EU data and non-discrimination laws.
11:40-12:00
Presentation: A03-3: FEDERATED LEARNING AND KNOWLEDGE DISTILLATION FOR AT-THE-EDGE EV MONITORING AND PREDICTIVE MAINTENANCE.
Pierluigi Dell’Acqua, Università degli Studi di Messina, Italy.
In recent years, deep neural networks have achieved success in Electric Vehicles monitoring, primarily due to their scalability with large-scale data and numerous model parameters. However, EVs use resource-constrained devices that struggle with complex models, and data privacy concerns prevent sharing data outside the owning device. Federated Learning and Knowledge Distillation have emerged as key solutions, enabling model simplification and distributed training on private data. FL allows training models in a distributed manner by keeping data locally on users’ devices rather than centralizing it on a server. This approach addresses privacy constraints but requires lightweight models deployable on edge devices.
To address this challenge, we propose an architectural solutions leveraging FL, KD, and model compression techniques such as quantization and pruning to create simplified Artificial Neural Networks suitable for deployment on resource-constrained devices. The proposed architecture integrates these compression techniques into a federated environment, supporting distributed training while ensuring that the models remain computationally feasible for EV monitoring and predictive maintenance applications. By combining FL, KD, and model compression, our approach aims to enable efficient and privacy-preserving Machine Learning model integration into Internet of Things devices within EVs, enhancing monitoring and maintenance capabilities in resource-limited settings.
Speaker:Pierluigi Dell’Acqua, Università degli Studi di Messina, Italy.
Pierluigi Dell’Acqua was born in Messina, Italy. He received the bachelor degree in Computer and Telecommunications Engineering and master degree in computer engineering from the University of Messina, in 2009 and in 2012 respectively. After then 10 years of experience as Software Engineering and Application Manager, he is now a PhD. student in the Department of Mathematics and Computer Sciences, Physical Sciences and Earth Sciences at University of Messina.
Presentation Link
Presentation: B03-3: TRUSTWORTHY EDGE AI SOLUTIONS IN AUTONOMOUS VEHICLE PERCEPTION AND SENSOR FUSION.
Ovidiu Vermesan, SINTEF, Norway.
The presentation explores edge AI’s critical role in enhancing autonomous vehicles’ safety and reliability. It discusses how edge AI processes data from multiple sensors in real-time, enabling quick and accurate perception of the vehicle’s surroundings. Emphasising trustworthiness, the presentation will address challenges such as data integrity, bias mitigation, and system robustness. By integrating edge AI practices and advanced algorithms, we can create autonomous systems that perform effectively and gain public trust, ensuring safer mobility for transporting goods.
Speaker: Ovidiu Vermesan, SINTEF, Norway.
Ovidiu Vermesan is Chef Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies and applying the developments to applications such as green mobility, energy, buildings, autonomous systems, electric connected, autonomous, and shared vehicles, and industrial manufacturing. He is currently the coordinator of the Edge AI Technologies for Optimised Performance Embedded Processing (EdgeAI) project.
12:00-12:20
Presentation: A03-4: PIONEERING THE HYBRIDIZATION OF FEDERATED LEARNING IN HUMAN ACTIVITY RECOGNITION.
Federico Montori, University of Bologna, Italy.
This study introduces a novel system, namely Hybrid Federated Learning (HFL), which combines two distinct learning techniques: Federated Learning (FL) and Centralized Learning (CL). FL enables the training of models on decentralized data sources, preserving data privacy and reducing communication overhead. In contrast, CL leverages data aggregation in a centralized manner to improve model performance. HFL capitalizes on the strengths of both approaches, aiming to overcome their individual limitations. This research examines the architecture and implementation of HFL, focusing on its potential to enhance model robustness, scalability, and privacy preservation. By integrating FL’s distributed learning paradigm with CL’s centralized model updates, HFL balances local data autonomy and global model optimization. Furthermore, we analyze the trade-offs and considerations associated with deploying HFL in a Human Activity Recognition (HAR) use case.
Speaker:Federico Montori, University of Bologna, Italy.
Federico Montori received the B.S. and M.S. degrees (summa cum laude) in computer science and the Ph.D. degree in computer science and engineering from the University of Bologna, Italy, in 2012, 2015, and 2019, respectively. He was a Visiting Researcher at Swinburne University of Technology (Australia), Luleå Tekniska Universitet (Sweden), and Technische Universitat Ilmenau (Germany). He is currently a Senior Assistant Professor at the University of Bologna. He participated in several EU projects and was WP Leader for the H2020 Project Arrowhead Tools. His primary research interests include mobile crowdsensing (MCS), pervasive and mobile computing, IoT automation, and data analysis for IoT scenarios.
Presentation Link
Presentation: B03-4:
12:30-13:30
Networking Lunch (Room Terrazza)
13:30-15:00
Session C03 (Room Stampace): EDGE AI SYSTEMS APPLICATIONS.
Session Co-chairs:
Silvester Sabathiel, Silicon Austria Labs, Austria; Ihsen Alouani, CNRS, Université Polytechnique Hauts-de-France.
The session showcases the transformative impact of edge AI across various industries, highlighting practical implementations and real-world successes. Participants will explore diverse applications, demonstrating how edge AI enhances decision-making, reduces latency, and optimises resource utilisation. The session aims to inspire attendees with innovative ideas and strategies for integrating edge AI into their applications, driving efficiency and growth.
Session D03 (Room Castello): SUSTAINABLE EDGE AI TECHNOLOGIES AND LIFECYCLE DEVELOPMENT.
Session Co-chairs:
Michael Cernusca, ams OSRAM AG; Modris Greitans, Institute of Electronics and Computer Science, Latvia.
The session focuses on integrating sustainability into edge AI systems’ design, deployment, and maintenance. Participants will explore eco-friendly practices and technologies that minimise energy consumption, reduce hardware waste, and enhance the longevity of edge devices. Speakers will discuss methodologies for evaluating the environmental impact throughout the AI lifecycle, from development to end-of-life. Implementations of sustainable practices to adopt greener strategies in edge AI projects will be presented.
13:30-13:50
Presentation: C03-1: MODEL SELECTION AND PROMPTING STRATEGIES IN RESOURCE-CONSTRAINED ENVIRONMENTS FOR LLM-BASED ROBOTIC SYSTEM.
Toms Eduards Zinars, Institute of Electronics and Computer Science, Latvia.
Large Language Models (LLMs) have become a major tool and asset in information processing and content generation. The performance of LLMs is tied to their size, with larger models being able to perform more complex tasks. However, these larger models require massive computing resources, which in some applications may not be available. Coupled with the diverse availability of model weights, quantization, and fine-tuning, it is desirable to find the best-performing LLM within a certain memory budget. Prompting strategy plays a similarly major role in the quality of generation, with various methods existing that attempt to induce desired behaviour. This is important for systems such as a high-level planner (HLP) for robotic control tasks, where maximum performance for the lowest memory cost is necessary for optimal operational capabilities, as better results from the model allow for more complex planning operations to be entrusted to the system. Under ideal circumstances, it would be possible to create a system that allows for the switching out of LLMs as newer and more performant weights become available. However, the practical reality is that minor changes in the contents of the prompt can benefit some models while disadvantaging others, making it harder to compare different models and making switching out models without testing irresponsible, as performance cannot be predicted with absolute certainty. This necessitates the development of robust testing and validation mechanisms for both new models and changes in the prompt.
Speaker: Toms Eduards Zinars, Institute of Electronics and Computer Science, Latvia.
Toms Eduards Zinars graduated from Riga Technical University with a Bachelor’s degree in “Mechatronics”. Continuing studies in RTU, in 2024 earning a master’s degree under the “Intelligent Robotic Systems” program. Since mid-2023 has worked at Institute of Electronics and Computer Science with a focus on high level planning for robotic systems. Interests include robotics, high-level planning, large language models.
Presentation Link
Presentation: D03-1: AI SUSTAINABLE APPLICATIONS – EXPERIENCING DRIVING.
Reiner John, AVL, Austria..
The vision of “Experiencing Driving” represents a groundbreaking transformation in vehicle technology, where the driving experience becomes truly immersive through the integration of cutting-edge technologies. At the heart of this vision is the concept of a “borderless view,” where displays replace traditionally opaque areas like the A- and B-pillars and the roof, supported by external cameras, to give the driver an uninterrupted 360-degree view. This immersive experience goes far beyond conventional systems: the driver is fully integrated into their surroundings, enhancing safety, comfort, and control. The seamless interaction between driver, vehicle, and environment creates a stress-free and more relaxed driving experience. Immersion in this context means that through Virtual Reality (VR) and Edge Computing, the driver experiences an environment that adapts to their individual needs, allowing them to perceive all aspects of driving in real-time, without limitations. Advanced sensors, AI-powered systems, and innovative displays turn the car into an extension of the driver, providing complete control and a more intense driving sensation. The synergy of VR, AI, and virtualization technologies redefines the driving experience, with innovations like cutting-edge headlights and 360-degree displays being core components of this vision.
Speaker: Reiner John, AVL, Austria.
Reiner John, an accomplished electrical engineer, began his career at Siemens Semiconductor Group before transitioning to Siemens Automation Group, where he specialized in embedded control development tools. His expertise expanded into knowledge-based embedded systems and automotive systems research at Siemens’ Central Research and Development. He later held management positions at Infineon Technologies and established their Silicon Foundry Taiwan Office. Since 2007, Reiner has been instrumental in initiating and coordinating publicly funded projects. Currently, as the Coordinator of Research Funding in Corporate Strategy at AVL Graz, he focuses on advancing electromobility and driving innovation in cutting-edge technologies and AI-driven digitalization.
13:50-14:10
Presentation: C03-2: FROM HUMAN EXPERT TO DIGITAL KNOWLEDGE: APPLYING OF METHODS FOR KNOWLEDGE EXTRACTION.
Silke Braun, Infineon, Austria.
The digitalization of domain knowledge from experts is crucial for capturing and preserving valuable knowledge and experience. It enables the creation of standardized processes and procedures, supports decision-making and problem-solving, and drives innovation and improvement. This talk gives an insight into the methods used to extract human knowledge within the development of the Smart Semiconductor Manufacturing Assistant (SSMA).
Presenter: Silke Braun, Infineon, Austria.
Silke Braun completed her studies in microsystems engineering at Mittweida University of Applied Sciences (Germany). In 2017, she started working at Infineon Technologies Dresden as a defect density and metrology engineer; since 2018 in Villach (Austria). She strengthened her knowledge in the field of quality management. Since 2022, she focuses her work on the field of funding and works in various EU-funded projects.
Presentation Link
Presentation: D03-2: ARTIFICIAL INTELLIGENCE IN MANUFACTURING LEADING TO SUSTAINABILITY AND INDUSTRY5.0.
Thomas Gutt, Infineon Technologies AG, Germany; Markus Tauber, Research Studio Austria.
The industrial world is rapidly moving towards its 5th industrial revolution (a.k.a. “Industry5.0”). Industry5.0 is reflected on the digital sovereignty in comprehensively sustainable production, through adopting, extending and implementing AI-enabled hardware, as well as AI tools & methods and semiconductors technology across the whole industrial value chain. It is expected that in doing so, manufacturing costs will be decreased, while at the same time, product quality will be increased through AI-enabled innovation, time-to-market will be shortened, and user acceptance of versatile technology offerings will be achieved and global supply chains tabilized. The above will, in turn, foster a sustainable development, in an economical, ecological and societal sense and act as enablers for the Green Deal. This paper accordingly discusses on the fundamental research areas relevant to further advancing the digitalized industry, its research and development achievements so far, as well as the challenges confronted, to boost industrial competitiveness through interdisciplinary innovations, establishing sustainable value chains and therefore contribute to the Digital Sovereignty.
Speaker: Thomas Gutt, Infineon Technologies AG, Germany; Markus Tauber, Research Studio Austria.
Thomas Gutt is a director of Funding Projects & Coordination at Infineon Technologies. He holds a dipl. phys. in nuclear physics from the university of Bochum. In more than 25 years of experience in the semiconductor industry he has worked as single process development engineer and as project manager for Semiconductor power switches (IGBT). Since 2017 he is coordinating the funding projects for industrial digitalization (like productive4.0 with 109 Partners).
Markus Tauber works as Chief Scientific Officer at Research Studios Austria, where he focuses on core digital technologies in the area of Edge-AI, Cyber Security and Autonomous Cyber Physical Systems and serves in the scientific council of the Inside Industry Association. In addition to this Markus is a Guest Professor at the Department of Telecommunications and Media Informatics, Faculty of Electrical Engineering and Computer Science, Budapest University of Technology and Economics. He brings many years of experience, e.g.as Researcher at the Austrian Institute of Technology (AIT), or Professor at the University of Applied Sciences Burgenland. Markus received a PhD in computer science with a focus on autonomic management in distributed storage systems from the University of St Andrews.
14:10-14:30
Presentation: C03-3: IMPLEMENTATION OF AN AIOT TEACHING KIT BASED ON LORA TECHNOLOGY AND EDGE AI.
Louis Noaille, Technext, France.
The link between education and new technologies has become increasingly important over the last few decades. Within the EdgeAI project, we are seeking to provide an educational tool, inspired by industrial issues, enabling teachers to introduce the various notions of AIoT. By making didactic use of the technical and technological advances made within Value Chain 3, we are designing an educational kit based on LoRa communication technology and using AI implanted directly into the sensors. The aim is to be able to follow the information as it travels along the communication network, from acquisition of the raw data to their display, including processing by the AI at the edge. This teaching kit will contain all the components needed to set up an intelligent communication network, and will give students the opportunity to address a large number of curriculum points at several levels (from high school to engineering school). It will also provide a first overview of the contribution that AIoT can make to sustainable development issues.
Speaker: Louis Noaille, Technext, France.
Louis Noaille is a Research and Development engineer at the Technext company, working on issues related to telecommunications and embedded AI.
Presentation Link
Presentation: D03-3: EXPLAINABLE AI LIFECYCLE.
Dinu Purice, Cognition Factory, Germany.
Explainable AI (XAI) Lifecycle encompasses the end-to-end development process of AI-based applications taking into account the need for transparency and interpretability. This talk highlights methodologies aimed at fostering trust, regulatory compliance, and predictability in industrial applications.
Speaker: Dinu Purice, CognitionFactory, Germany.
Dinu Purice is a Solutions Architect, have Masters in Physics/Medical Imaging, and the Area of expertise is Computer Vision.
14:30-14:50
Presentation: C03-4: CONSCIOUS AGENTS INTERACTION FRAMEWORK FOR INDUSTRIAL AUTOMATION.
Polina Ovsiannikova, Aalto University, Finland.
This work-in-progress paper explores the integration of human cognition and interaction models into industrial automation systems. We begin by examining how human cognitive patterns can be applied to the development of conscious agents within these systems. From an interaction standpoint, we analyze the dual role of humans as both users and workers within automated environments. The convergence of these perspectives enables the creation of intelligent, multi-agent systems where humans function as equal agents. Such systems are characterized by true flexibility, as each component can independently assess its capabilities and collaboratively plan actions to achieve both collective (external) and individual (internal) goals. We present two case studies to illustrate these concepts: the first involves the automation of a cruise ship’s HVAC system, where agents (cabin units) negotiate a temperature setpoint based on human behaviour (user-focused scenario). The second case study examines a distributed vertical farm, where modules must coordinate their energy consumption, demonstrating steps towards cognitive reasoning in machines. These examples underscore the potential of integrating human-like reasoning and interaction into automated systems, advancing their adaptability and efficiencymmW.
Speaker: Polina Ovsiannikova, Aalto University, Finland.
Polina Ovsiannikova, eceived B.Sc. degree in software engineering, M.Sc. degree in applied mathematics and computer science from ITMO University, Saint Petersburg, Russia and Ph.D. from Aalto University, Finland, Espoo. Her research interests include energy consumption optimization and application of human cognition to industrial automation, both within a muti-agent systems paradigm.
Presentation Link
Presentation: D03-4: AI BASED ATTACK DETECTION ON RESOURCE CONSTRAINT SECURE MICRO EDGE DEVICES.
Ujunwa Irogbarachi NXP Semiconductors, Germany.
Nowadays, secure elements are part of everyday life, as these devices store and process highly sensitive data. Typical applications of secure element products can be seen in smartcards, smartphones, wearables, and accessories across the payment, transit, governmental, and hospitality sectors. As these secure elements are integrated into many systems where the devices are constantly in operation and exposed to ever-increasing environmental stress, simple sensor reset configurations implemented in the secure elements may not be sufficient.
To ensure the correct functionality of these resource-constrained devices without negatively impacting the user experience and compromising the security of user data, algorithms based on anomaly detection techniques using artificial intelligence becomes necessary for intelligent identification of attacks, differentiating between false alarms and real threats to reduce the number of unnecessary resets and achieve low threat miss rates.
Speaker: Ujunwa Irogbarachi, NXP Semiconductors, Germany.
Ujunwa Irogbarachi is a Senior Systems Engineer for Secure Element products working with the Secure Connected Edge business unit at NXP semiconductors, Hamburg. She holds a master’s degree in information and communications systems from the Hamburg University of Technology (TUHH), specializing in secure and dependable IT systems. Her bachelor’s degree is in Computer Science and Economics from the Obafemi Awolowo University, in Nigeria.
15:00-15:30
Afternoon Coffee/Tea Break (Room Terrazza)
15:30-17:00
Edge AI Cluster Internal Meeting (Room Castello)
17:00
End
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title – Room A01
Architecting Edge AI Workflows for Predictive Maintenance in Industrial Applications
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation A01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation A01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation A01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
| Networking Lunch |
Parallel Session C01: Title – Room C01
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation C01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation C01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation C01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation C01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title – Room E01
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation E01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation E01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation E01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation E01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
| End |
| Dinner |
Parallel Session B01: Title – Room B01
Architecting Edge AI Workflows for Predictive Maintenance in Industrial Applications
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation B01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation B01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation B01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Parallel Session D01: Title – Room D01
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation D01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation D01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation D01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation D01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Parallel Session F01: Title – Room F01
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation F01_1 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation F01_2 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation F01_3 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Presentation F01_4 – Name Presentert, Company Affiliation, Country
Your content goes here. Edit or remove this text inline or in the module Content settings. You can also style every aspect of this content in the module Design settings and even apply custom CSS to this text in the module Advanced settings.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
9:00-17:00
Poster Session (Room Terrazza): 21-23 October, 2024
Poster: P01 : ENABLING RELIABLE COMMUNICATION IN V2X SCENARIOS.
Anja Dakić (1), Benjamin Rainer (1), Thomas Zemen (1), Stefan Jakšić (1), Anamarija Momic (2), Peter Priller (2), Xiaochum Ye3, Guo Nan (3).
(1) AIT Austrian Institute of Technology GmbH, Center for Digital Safety & Security, Giefinggasse 4, 1210 Vienna, Austria; (2) AVL List GmbH, Hans-List-Platz 1, 8020 Graz, Austria; (3) Institute of Computing Technology, Chinese Academy of Sciences, No. 6 Kexueyuan South Road Zhongguancun, Haidan District Bejing, China.
Poster: P02: CONTINUOUS-TIME NEUROMORPHIC PROCESSING AT THE SENSOR EDGE.
Amir Zjajo, Sumeet Kumar, Innatera Nanosystems B.V., The Netherlands.
Poster: P03: EDGE AI APPLICATIONS IN DIGITAL INDUSTRY.
Silke Braun, Josef-Anton Moser, Infineon, Austria.
Poster: P04: CONTINUOUS OPTIMIZATION AND BENCHMARKING OF AI MODELS ON MICROCONTROLLERS: A TINYMLOPS FRAMEWORK FOR REAL-WORLD APPLICATIONS.
Mattia Antonini, Massimo Vecchio, Fabio Antonelli, Fondazione Bruno Kessler, Trento, Italy.
Poster: P05: EDGE AI APPLICATIONS IN AGRI-FOOD AND BEVERAGE.
Marcello Coppola, STMicroelectronics, France.
Poster: P06: EDGE AI APPLICATIONS IN ENERGY INDUSTRY.
Ovidiu Vermesan, SINTEF AS, Norway.
Poster: P07: EDGE AI APPLICATIONS IN DIGITAL SOCIETY.
Gerardo Daalderop, NXP Semiconductors, the Netherlands.
Poster: P08: EDGE AI APPLICATIONS IN MOBILITY INDUSTRY.
Janis Arents, EDI, Latvia.
Poster: P09: CHIPS JU EDGEAI PROJECT.
Athanasia Thraskia, ITML, Greece.
Poster: P10: NON-VOLATILE MEMORY (NVM) BASED IN-MEMORY COMPUTING (IMC) SOLUTIONS OF ANDANTE PROJECT.
Alexandre Valentian, CEA, France, Mario Diaz Nava, STMicroelectronics, France.
Poster: P11: FLEXIBLE DNN MAPPING MODEL FOR NOC-BASED DATAFLOW ARCHITECTURES.
William Fabre, Hana Krichene, Rohit Prasad, and Caaliph Andriamisaina, CEA, France.
Poster: P12: APPLICATION OF DEEP NEURAL NETWORKS FOR FUNDAMENTAL FREQUENCY EXTRACTION IN OCEAN CURRENT METERS USING EDGE COMPUTING ON ULTRA LOW POWER MICROCONTROLLERS.
Juan Montiel-Caminos, Nieves G. Hernandez-Gonzalez and Javier-Sosa. University of Las Palmas de Gran Canaria, Spain.
Poster: P13: COMPUTER VISION AT THE EDGE.
Arno van der Weijden, Technolution BV, The Netherlands.
Poster: P14: NEW MODULAR BUILDING BLOCKS APPROACHES FOR CLASSIFICATION IN REAL WORLD APPLICATIONS IN EDGE AI.
Lucas Mohimont, Université de Reims Champagne-Ardenne (URCA), France.
17:00
End
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session I: Edge AI Technology Advancements
9:00-9:30
Keynote presentation I_1
9:30-10:00
Keynote presentation I_2
10:00-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A01: Edge AI Tools and Methods.
Session Co-chairs:
Fabian Chersi, CEA-List, France, Frédéric Pétrot, Univ. Grenoble Alpes, France
This session is dedicated to the presentation of tools and methods aimed at simplifying and accelerating the design flow, from application development to deployment, from neural network exploration, learning, quantizing, testing, to the generation of optimized hardware-aware code and deployment on edge devices constrained by power consumption or dissipation, latency, form factor and/or cost criteria.
Keywords: edge devices, quantization, pruning, NN optimization, NN deployment, tools for edge devices.
Session B01: Edge AI Machine Vision.
Session Co-chairs:
Alain Pagani, DFKI, Germany, Sadique Sheik, SynSense AI, Switzerland
The session covers presentation addressing edge AI machine vision technology and applications developments based on different algorithms and techniques. The presenters focus on the optimisation techniques and AI accelerators capable of processing inference at minimal power consumption. The implementations include ML algorithms and neural networks used in machine vision tasks optimises for edge devices with constrained resources.
Keywords: edge AI, machine vision, machine learning, deep learning, image signal processors, convolutional neural networks.
11:00-11:20
Presentation: A01_1: Optimization and Compression of Computer Vision and Generative AI Models for Neuromorphic Accelerators.
Tim Llewellynn, Bonseyes Community Association, Switzerland.
In many AI applications, models typically demand extensive training datasets and benchmark data for robust performance. However, gathering such vast and diverse datasets can be especially challenging in edge computing environments. This presentation addresses this critical challenge by showcasing how Generative AI and Computer Vision foundational models can help mitigate the burden by generating synthetic data and benchmarks that closely resemble real-world scenarios. Applicability to Neuromorphic model development will be discussed and AI development in edge computing, where data scarcity is a common constraint.
Speaker: Tim Llewellynn, Bonseyes Community Association, Switzerland.
Tim Llewellynn is President of the Bonseyes Community Association and has 20 years in high-tech startups and corporate R&D in data-driven software development for cloud, mobile and edge computing applications. At Bonseyes Community Association (BCA), is a non-profit AI-focused association, he drives the democratization of artificial intelligence for the benefit of all society, ensuring that AI, from narrow to general, creates highly autonomous systems that benefit everyone, not just a select few. Tim received his B.E. and M.E. (Hons) degrees in Electrical Engineering from the University of Canterbury, New Zealand in 1999, and 2001.
Presentation: B01_1: Image Signal Processor (ISP)
Tunning Machine Learning (ML) methods.
Sepehr Bijani, NXP Semiconductors, Munich, Germany
Image Signal Processor (ISP) is responsible for improving camera signal quality and producing a high-quality image. ISP has a vast number of parameters which should be tunned based on both camera sensor and the environment which camera is operating. Sensor related ISP parameters should be tunned offline for each camera sensor. Parameters related to environment such as white balancing gains should be fine-tuned in the field during the runtime. The offline phase is cumbersome and costly. At the same time, the runtime for fine-tuning should be fast and accurate. Therefore, the Tuner needs to achieve two goals: a) tune the sensor related parameters automatically in lab and b) create a data set for different scenario in lab and train a Machine Learning (ML) runtime for fine-tunning ISP in the field. We first create a solver for finding static camera related parameters. Then, we create the dataset of camera scenes and optimal ISP parameters found by the same solver for each scene for training runtime ML model. The trained ML model will inference optimal ISP parameters for the scene in the field.
Speaker: Sepehr Bijani, NXP Semiconductors, Munich, Germany.
Sepehr is SW engineer working for NXP Germany. His main research topic is AI/ML in image processing applications. His professional carrier focuses on embedded SW engineering.
11:20-11:40
Presentation: A01_2: Shrinking Embedded Machine Learning Applications: FMCW Radar Gesture Sensing on Cortex-M0+ under 10KB and 10mW.
Jonas Daugalas, Infineon Technologies AG, Germany
AI-enabled millimeter-wave radar sensors are becoming increasingly popular solutions for human-machine interaction tasks. However, due to compute and memory requirements, signal processing and machine learning approaches pose challenges for deploying such solutions to low-cost and low-power edge devices. In this work we demonstrate signal processing optimization techniques and neural network inference engine trade-offs that allow us to deploy a 60GHz FMCW radar-based gesture recognition solution on an Infineon PSoC™ 4 microcontroller equipped with an Arm® Cortex-M0+ processor. Our proposed system robustly recognizes five distinct hand gestures, requires less than 10 milliwatts of power, 9.7KB of RAM, 107KB of EPROM, 73% of processor time, and with microTVM, among other three evaluated inference engines, achieves the best trade-off between manual effort and in-field performance.
Speaker: Jonas Daugalas, Infineon Technologies AG, Germany.
Jonas Daugalas has a background in Computer Science, bachelor’s degree in CS from Vilnius University and master’s degree in CS from Technical University of Munich. Software engineering experience in the fields of particle physics, computer-aided design, physical simulation, machine learning and digital signal processing.
Jianyu Zhao has background in Electrical and Information Engineering: received a B. Eng. from Tianjin University in 2015 and a M. Sc. from Technical University of Munich in 2018. Work experience on algorithm design and deployment for smart sensing technologies at Infineon Technologies AG since 2018. Current research interest: neural network embedding on edge devices.
Presentation: B01_2: Ultra-low Power Car Plate Recognition at the Edge.
Dávid Isztl, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
So far, only a few edge devices include AI accelerators that are capable of processing inference at minimal power consumption. As an example, the MAX78002 provides a 2Mbytes CNN engine and an Arm Cortex-M4 with 384KB of SRAM. On one hand, the device can work battery-powered for up to 3 years. On the other hand, severe compromises and complex procedures are necessary to adapt AI networks for specific applications. Using license plate recognition as an example, the talk will show how the shortcomings of the small memory footprint and low image resolution offered by the AI accelerator can be overcome.
Speaker: Dávid Isztl, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
Dávid Isztl is a research assistant at the Institute of Embedded Systems at ZHAW. Currently, his research is focused on high-performance embedded platforms, including AI accelerators and Nvidia NX/AGX platforms. Currently, he is focusing on low-power AI processing on the edge.
11:40-12:00
Presentation: A01_3: Aidge: A New Framework for Deep Neural Network development, Training, and Deployment on the Edge.
Fabian Chersi. CEA-List, FRANCE
Aidge is a comprehensive framework for fast and accurate Deep Neural Network (DNN) building, training, testing and deployment. The platform integrates tools for database construction, data pre-processing, network building, benchmarking, and hardware export to various targets. It is particularly useful for DNN design and exploration, allowing simple and fast prototyping of DNN with different topologies. Export hardware targets include CPU, DSP, and GPU as well as custom hardware code generation with High-Level Synthesis for FPGA and dedicated configurable DNN accelerator IP.
Speaker:Fabian Chersi. CEA-List, FRANCE.
Dr. Fabian Chersi has a BSC and a Master in Electronic Physics, and a PhD in Computational Neuroscience and Robotics. After his studies, he initially did a postdoc at the Institute for Neuroinformatics in Zurich and at UCL in London, then he moved to Paris where he worked in private companies on the development of neuromorphic algorithms and ASICs for deep neural networks. He currently works at the Alternative Energies Commission (CEA) in Paris as a senior AI architect on low power AI chips for the edge.
Presentation: B01_3: Designing Lightweight CNN for Images: Architectural Components and Techniques.
Lilian Hollard, Université de Reims Champagne-Ardenne, France
While neural networks have brought about impressive advancements in computer vision tasks, these achievements heavily depend on computationally demanding resources, restricting their deployment. The presentation unveils the landscape of deep learning architecture optimization tailored for mobile and Edge devices.
Speaker: Lilian Hollard, Université de Reims Champagne-Ardenne, France.
First year Ph.D. student at LICIIS from the University of Reims Champagne Ardenne, development of Artificial Intelligence models for viticulture: Disease identification and yield estimation. Adaptation and optimization of Deep Learning models for Edge and mobile devices.
12:00-12:20
Presentation: A01_4: A Low Power And High-Performance Software Approach to Artificial Intelligence On-Board.
Pablo Ghiglino, Klepsydra Technologies, Switzerland.
Future spacecraft require increased autonomy, more sensors, and better processing power for space exploration, Earth Observation, and robotics. These advancements bring challenges like low determinism, high power consumption, data losses, and response delays. This article presents a novel approach to on-board artificial intelligence (AI) using data pipeline and parallelization techniques to enhance deep neural networks (DNN) in space applications. The results show significantly improved data processing speed and power efficiency, enabling practical AI in space.
Speaker:Pablo Ghiglino, Klepsydra Technologies, Switzerland.
Dr Pablo Ghiglino, Klepsydra founder and CEO. Over 10 years of experience in the finance industry, regarded as a highly skilled software developer. Parallel to his work in finance, Pablo successfully studied for a PhD at the Surrey Space Center in Aerospace Engineering in the field of “High Precision Planetary Landing”. After his studies, he got involved in several aerospace projects, where he seized the opportunity and created Klepsydra to respond to the needs for cutting edge computing by implementing successful software techniques from the highly competitive financial sector in embedded software frameworks for its use in other industries.
Presentation: B01_4: Upstream Mask Tuning of the Plain Vision Transformer for Domain Generalization on Downstream Segmentation Tasks.
Tommie Kerssies, Eindhoven Technical Univerity, the Netherlands.
Deploying perception models on edge devices requires robustness to real-world distribution shifts. We reveal that a fine-tuned plain vision transformer excels in out-of-distribution generalization but is hindered by single-scale features and image-level, as opposed to pixel-level, pretraining. To mitigate this, we employ a mask-tuning step on a large, high-resolution dataset consisting solely of masks, guided by a frozen teacher for semantic context. Our work aims to redirect the domain generalization field towards optimizing pretraining rather than complicating fine-tuning.
Speaker: Tommie Kerssies, Eindhoven Technical Univerity, the Netherlands.
Tommie Kerssies is a PhD candidate in the Mobile Perception Systems Lab at Eindhoven University of Technology, under the guidance of Gijs Dubbelman. Specializing in the generalization capabilities of foundation models, Tommie brings a strong academic background with a master’s in data science and a bachelor’s in computer science.
12:30-13:30
Networking Lunch
13:30-15:30
Session C01: RISC-V-based Edge AI Hardware Acceleration.
Session Co-chairs:
Konstantinos Georgopoulos, Iakovos Mavroidis, Technical University of Crete, Greece
RISC-V-based edge devices are rapidly expanding their scope and application range. They are efficient in executing a variety of AI algorithms close to the application environment, ensuring energy efficiency while maintaining high performance. Furthermore, the open-source architecture of RISC-V enables researchers to customize and enhance it, incorporating low-power AI and security accelerators tailored to the specific needs of edge applications. This session will focus on RISC-V-based edge architectures and their diverse applications.”
Keywords: RISC-V, hardware accelerator, reconfigurable hardware, AI accelerator, HW/SW co-design, energy consumption, HW-based security, ΙοΤ.
Session D01: Edge AI Accelerators Architectures.
Session Co-chairs:
Mario Diaz Nava, STMicroelectronics, France, María Loreto Mateu Sáez, Fraunhofer/IIS, Germany.
Efficient Edge AI system solutions (with highly energy efficient, low latency and high integration) require new HW Processors that support AI advanced features based on neuromorphic architectures implemented in different embedded non-volatile memory and SRAM technologies. This session presents a selection of different AI accelerators capable of addressing different areas of Edge applications. These hardware processors are essential for optimal Edge solutions, but a trade-off between cost and technological features is necessary.
Keywords: HW accelerators, eNVM, FPGA, architecture, neural networks, edge IoT devices, framework, toolchain, dataflow, flexibility, scalability.
13:30-13:50
Presentation: C01_1: How to Accelerate Your Application on RISC-V Using a Coprocessor Interface.
Dr. Christian Herber, NXP, Germany.
RISC-V is becoming the most popular ISA for advanced applications like AI due to its modularity and customizability. However, customization comes with several downsides, like ecosystem fragmentation and design complexity. A clearly defined coprocessor interface can reduce design efforts and improve interoperability. This presentation outlines the use-cases and challenges in a coprocessor interface based on concrete interface definitions.
Speaker: Dr. Christian Herber, NXP, Germany.
Christian Herber the European Principal RISC-V Architect at NXP. Based in Hamburg, his work focuses on innovation management and advanced IP for RISC-V processors in Europe. He has a background in automotive processing, networking, and software. He holds a PhD in Electrical Engineering and an M.Sc. in Electrical Engineering and Information Technology, all from Technical University of Munich, Germany.
Presentation: D01_1: Efficient Mapping of Neural Networks on FPGA With Generated VHDL code.
Thierry Delafontaine, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
FPGAs are an ideal platform to accelerate AI processing at the edge. However, straightforward toolchains to deploy AI accelerators without in-depth knowledge of FPGA programming are lacking. We present a framework that allows the design of custom AI accelerators for FPGAs by generating application specific VHDL code. The framework proposes a dataflow-style hardware mapping that is specific to the NN and optimally uses the available resources. We focused on providing an easy-to-use interface for users and an expandable core that allows developers to add custom operators, optimizers, objectives, metrics, and callbacks.
Speaker: Thierry Delafontaine, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
Thierry Delafontaine is a researcher at the Institute of Embedded Systems at the University of Applied Sciences in Zürich. His research topics include embedded AI and accelerating Neural Networks on embedded devices. Currently, he is building the iAIa framework, that he presents at the EEAI conference.
13:50-14:10
Presentation: C01_2: The Beauty of RISC-V Customization for Edge AI HW Acceleration.
Arturo Prieto, Lund University, Sweden.
Edge AI is becoming an increasingly popular topic due to the continuous development of AI applications avoiding the dependencies on-the-cloud. The use of sensitive data or latency constraints push for processing on edge devices. Battery-limited and area-constrained devices offer the opportunity for hardware designers to find efficient architecture solutions. In this presentation, a RISC-V platform is presented as the conductor for the customized integration of hardware accelerators.
Speaker: Arturo Prieto, Lund University, Sweden.
Arturo Prieto received his B. Sc. degree in Electronics Engineering in 2017 from the Polytechnic University of Valencia, Spain, and the M. Sc. degree in Electronic Design in 2019 from Lund University, Sweden. He is currently a PhD student in the Digital ASIC group at the department of Electrical and Information Technology (EIT), Lund University. His doctoral studies focus on beyond-von Neumann architectures for edge computing.
Presentation: D01_2: A Scalable and Flexible Interconnect-based Dataflow Architecture for Edge AI Inference.
Hana KRICHENE. CEA-List, France.
The scalable approach of edge AI inference has increased its computational complexity due to the involvement of multiple properties. This work proposes an interconnect-based dataflow architecture based on many Processing Elements to overcome such problems. The proposed architecture can efficiently handle different convolutions featuring different shapes of input image/feature maps and filters, with data reuse and communication-computation overlap. It is scalable and configurable to adapt to different CNN layers.
Speaker: Hana KRICHENE. CEA-List, France.
Dr. Hana Krichene obtained her Ph.D. Degree in Computer Science in 2015 from the University of Lille. She joined CEA LIST (Saclay, France) in 2017 to participate in the development of the Future E/E Architecture for Automotive Computing Environment. Since 2020, she conducts study research on NoC optimized for AI hardware accelerators of future systems used by the industrial partners of the project. Her current research focuses on developing an advanced dataflow architecture and an environment for mapping and estimating the cost of running a DNN application on a given dataflow architecture.
14:10-14:30
Presentation: C01_3: Heterogeneous SoA Processing on the Edge.
Yannis Papaefstathiou, Aristotle University of Thessaloniki, Greece.
The REBECCA project is developing efficient edge-AI systems that can overcome physical limitations of edge devices and meet regulations and constraints in numerous distinct application domains. This involves designing and developing hardware, software, and middleware that can accelerate computation-intensive parts of both AI and conventional applications, ensure deterministic response times, and satisfy safety constraints. The project is working towards enhancing European strategic autonomy and sovereignty by using an open, royalty-free CPU architecture (i.e., RISC-V), and address performance safety and security concerns associated with IoT and edge devices.
Speaker: Yannis Papaefstathiou, Aristotle University of Thessaloniki, Greece.
Yannis Papaefstathiou is a Professor at ECE School and Aristotle University of Thessaloniki and a co-founder and CEO of Exascale Performance Systems (EXAPSYS). His main research interests are in the design and implementation methodologies for CPS with tightly coupled design parameters and highly constrained resources as well as in heterogeneous High-Performance Computing (HPC) systems and the associated programming/development tools. He was granted a PhD in computer science at the University of Cambridge in 2001, an M.Sc. from Harvard University in 1997 and a B.Sc. from the University of Crete in 1996. He has published more than 100 papers in IEEE and ACM-sponsored journals and conferences. He has participated in more than 15 competitively funded research projects in Europe with a cumulative budget share of more than €9 million.
Presentation: D01_3: DeepEdgeSoC: End-to-end Deep Learning Framework for Edge IoT Devices.
Mhd Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
In the IoT domain, FPGAs are considered a powerful alternative to general purpose graphics processing units (GPGPUs) since their flexible architecture can run the DNNs with much less energy. The enormous amount of effort and time required for the entire end-to-end edge-aware deployment urged us to develop DeepEdgeSoc, an integrated framework for DL design and acceleration. DeepEdgeSoc is an overarching framework under which DNNs can be built. DeepGUI, a visual drag-and-drop DNN design environment, plays an important role in accelerating the network design phase. In DeepEdgeSoc, the networks can be quantized and compressed to suite the underlying edge devices in terms of size and energy.
Speaker: Mhd Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
After receiving his master degree in Embedded Computing Systems at the Technical University of Kaiserlslautern-Landau (RPTU), Mr. Mhd Rashed Al Koutayni joined the Augmented Vision department of the German Research Center for Artificial Intelligence (DFKI) in Kaiserslautern (Germany) as a scientific researcher while pursuing at the RPTU university as a PhD student. His main research focus is FPGA-Based Hardware Acceleration of Deep Neural Networks for Computer Vision.
14:30-14:50
Presentation: C01_4: Why Running Kubernetes on RISC-V is Important for AI Applications.
Antony Chazapis, FORTH, Greece.
Kubernetes is the leading open-source container orchestration platform, widely adopted for easily developing and running scalable applications from the Cloud to the Edge. And now, through K3s (a CNCF-certified distribution), one can also run Kubernetes on devices with processors that use the open RISC-V instruction set. In this talk we will present the process of supporting the new architecture and discuss the prospects this brings for AI applications in future Edge platforms.
Speaker: Antony Chazapis, FORTH, Greece.
Antony Chazapis is a postdoctoral researcher at the CARV Lab of FORTH-ICS. He works on distributed, large-scale platforms for both compute and storage that span the whole computing continuum, from the Cloud to the Edge. His experience and interests include acceleration technologies for Machine Learning/AI in both HPC and the Cloud, Cloud-native software architectures, bridging Cloud with HPC, as well as DevOps and MLOps practices.
Presentation: D01_4: A FeFET-based Non-Volatile Memory and AI Accelerator Macro.
Marko Noack, Ferroelectric Memory GmbH, Germany.
The presentation focuses on the design of a highly configurable AI accelerator macro for convolutional neural networks using ferroelectric NVM as weight storage and which serves as a building block for edge AI systems. We furthermore present a test setup in which the digital part of the circuit was implemented on an FPGA using weights stored on a separate FeFET NVM chip to demonstrate image processing tasks.
Speaker: Marko Noack, Ferroelectric Memory GmbH, Germany.
Marko Noack received the M.Sc. degree in electrical engineering from Technische Universität Dresden, Germany in 2011. He was working on VLSI implementations of neuromorphic circuits before he co-founded the Ferroelectric Memory GmbH (FMC) in 2016, where he is currently Member of Technical Staff and project leader. He works on non-volatile memory circuit designs based on ferroelectrics.
15:00-15:30
Afetrnoon Coffee/Tea Break
15:30-17:00
Session E01: Trustworthy Edge AI.
Session Co-chairs:
Patrick Pype, NXP Semiconductors, Belgium, Kay Bierzynski, Infineon Technologies AG, Germany.
Trustworthy edge AI addresses the development and deployment of AI systems at the edge of a network with a focus on ensuring system dependability e.g., reliability, security, privacy, and ethical behaviour among other properties. Ensuring trustworthiness in edge AI-based systems requires a multidisciplinary approach that combines expertise in electronics, sensing, IoT, AI, cybersecurity, privacy, ethics, and regulatory compliance. The session addresses the concepts for edge AI trustworthiness and the challenges for developing and deploying edge AI trustworthy systems.
Keywords: explainable edge AI, interpretable edge AI, transparency, edge AI bias mitigation, consent mechanisms, anonymisation, adverbial defence, accountability, verification, validation and testing.
Session F01: Optimisation Methods and Techniques for Neural Networks.
Session Co-chairs:
Gerardo Daalderop, NXP Semiconductors, the Netherlands, Alberto Faro, DEEPSENSING SRL, Italy.
Optimisation methods and techniques for neural networks depend on the specific problem, architecture, and available AI-based HW resources. Experimentation and tuning are required to find the optimal combination for a given neural network task. The session covers various optimisation methods and techniques used in training neural networks to find the set of model parameters that minimise a predefined loss functions.
Keywords: Bayesian neural networks, pruning techniques, energy, size, accuracy, hierarchical-controlling system, neural network processing.
15:30-15:50
Presentation: E01_1: Efficient AI-based Attack Detection Methods for Sensitive Edge Devices and Systems.
Celine Thermann, University of Lübeck, Germany.
Sensitive data is stored on an increasingly large number of edge devices, representing a lucrative target for attackers. Hardware and performance overhead, as well as environmental stresses, play a crucial role in the selection of appropriate solutions. In this research contribution, we are examining the requirements for Artificial intelligence (AI)-based mechanisms and the sensing capabilities of state-of-the-art secure devices.
Speaker: Celine Thermann, University of Lübeck, Germany.
Celine Thermann is a research assistant at the Institute of Computer Engineering at the University of Lübeck, Germany. She received both a Bachelor and Master degree in 2020 and 2022 in IT-Security from the University of Lübeck, respectively. Her main research interest is deep learning and embedded system security, and she focuses on the trustworthiness of AI accelerators.
Presentation: F01_1: Memristor Based Bayesian IMC.
Tifenn HIRTZLIN, CEA-LETI, France.
This presentation examines three methods employing resistive memory’s natural randomness for Bayesian near-memory and in-memory computing. The approaches encompass Bayesian machines optimized for near-memory computing, Bayesian neural networks that capitalize on synaptic randomness, and Bayesian learning using the Metropolis-Hastings Markov Chain Monte Carlo technique. These techniques yield competitive accuracy compared to traditional software methods while enabling decision uncertainty assessment.
Speaker: Tifenn HIRTZLIN, CEA-LETI, France.
Tifenn HIRTZLIN earned a Ph.D. in Nanosciences and Electrical Engineering from the University of Paris-Saclay, France, in 2020, and is currently a Research Engineer at CEA Leti. He is at the forefront of designing intelligent memories for the Internet of Things (IoT), drawing inspiration from bio-inspired concepts such as the Bayesian approach to brain function. Tifenn’s primary responsibility involves designing electrical circuits for application-specific integrated circuits (ASICs) that exploit non-volatile memory devices, and evaluating their performance, contributing significantly to the advancement of IoT technology.
15:50-16:10
Presentation: E01_2: Explainability and Interpretability Concepts for Edge AI Systems.
Angelo Genovese, Università degli Studi di Milano, Italy.
The presentation gives an overview of the recent techniques for AI explainability and interpretability, with a specific focus on Edge computing. The presentation will discuss the motivations for explainability and interpretability, a taxonomy of the latest methodologies, the most common benchmarking approaches, as well as challenges, open issues, and research trends.
Speaker: Angelo Genovese, Università degli Studi di Milano, Italy.
Angelo Genovese received the Ph.D. degree in computer Science from the Università degli Studi di Milano, Italy, in 2014. He has been an Associate Professor in Computer Science with the Università degli Studi di Milano since 2022. He has been a Visiting Researcher with the University of Toronto, Toronto, ON, Canada. Original results have been published in over 60 papers in international journals, proceedings of international conferences, books, and book chapters. His current research interests include signal and image processing, artificial intelligence, industrial and environmental monitoring systems.
Presentation: F01_2: Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices.
Duc V Le, University of Twente, The Netherlands.
Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This study analyses four pruning techniques for efficient reduction of machine learning model size, based on compression-accuracy balance. Evaluation includes impact on energy use during inference on a Raspberry Pi 4B using MobileNetV2. We examine trade-offs among energy, size, and accuracy, aiding engineers in selecting optimal pruning for embedded ML models. The research is expected to provide engineers with a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.
Speaker: Duc V Le, University of Twente, The Netherlands.
Dr. Le Viet Duc is Assistant Professor at the Department of Computer Science at the University of Twente. Inspired by the way the human brain learns the world, he is eager to develop sustainable AI models for that is robust to change in data distribution caused by new environments, bridging the gap between lab and word. His focus is on finding efficient methods to learn, transfer and distil knowledge hidden in different datasets, especially unlabelled ones. As he has 10 years R&D experience in developing smart and wearable and industrial devices, he is also interested in applying his research to industrial products for a sustainable world.
16:10-16:30
Presentation: E01_3: Approximate Computing in a Federated Learning Context: Gains Beyond Energy.
Ihsen Alouani, Université Polytechnique Hauts-De-France, France.
In energy-constrained systems at the edge, the adoption of federated learning has garnered significant attention due to its dual capacity to address privacy concerns while achieving accurate model training without resource-intensive methodologies. The Approximate Computing paradigm, characterized by leveraging controlled imprecision to achieve computational efficiency, has emerged as one of the solutions towards a trade-off between performance and energy consumption in fault-tolerant applications. Expanding on these foundational concepts, the presentation proposes an innovative approach, i.e., “Approximate Federated Learning.” Within this framework, nodes collaboratively undertake training and inference procedures utilizing approximate computing techniques, thereby accentuating energy efficiency gains. The study extends the benefits beyond energy conservation, highlighting substantial security enhancements achievable within this amalgamation. This holistic framework advances energy optimization in edge scenarios and bolsters the overall security and trustworthiness of edge AI systems.
Speaker: Ihsen Alouani, Université Polytechnique Hauts-De-France, France.
Ihsen Alouani received his master’s degree from the National Engineering School of Sousse, Tunisia, and his Ph.D. in 2016 from the Université Polytechnique Hauts-de-France. After his PhD, he was a Postdoc at Technical University of Munich, Germany followed by an Associate Professor at IEMN Laboratory (CNRS-8520), Université Polytechnique Hauts-de-France in 2018. He served as Guest Editor for several journals including Microprocessors & Microsystems and is a member of the Program Committee of several international conferences including Design Automation Conference (DAC) as well as Asia and South Pacific Design Automation Conference (ASP-DAC). His research interest includes machine learning security and privacy, Approximate Computing, and Embedded Systems reliability and security.
Presentation: F01_3: SENECA: Flexible and Scalable Neuromorphic Processor for Efficient Event-driven Neural Processing.
Guangzhi Tang, imec, the Netherlands.
The lack of flexibility in most neuromorphic architecture designs results in significant performance loss and inefficient memory usage when mapping various neural network algorithms. We present SENECA, a digital neuromorphic architecture that balances the trade-offs between flexibility and efficiency using a hierarchical-controlling system. A SENECA core contains two controllers, a flexible RISC-V-based controller, and an optimized controller (Loop Buffer). This flexible computational pipeline allows for deploying efficient mapping for various neural networks, on-device learning, and pre-post processing algorithms. The hierarchical-controlling system introduced in SENECA makes it one of the most efficient neuromorphic processors for event-driven neural network processing.
Speaker: Guangzhi Tang, imec, the Netherlands.
Guangzhi Tang is a researcher at imec based in Eindhoven, the Netherlands. His research focuses on neuromorphic computing, hardware-aware neural networks, and robotics. He has extensive experience working in the field of neuromorphic computing to explore the real-world advantages of digital neuromorphic architectures. He received his Ph.D. and M.Sc. degrees in Computer Science from Rutgers, the State University of New Jersey in the United States, and his B.Sc. degree in Computer Science from Nanjing University in China.
16:30-16:50
Presentation: E01_4: Shaping Edge AI That Flows With Humans: Designing Trustworthy and Useful AI Applications.
Peter Moertl, Virtual Vehicle Research GmbH, Austria.
The introduction of artificial intelligence solutions in industrial work processes leads in most cases to changes in human tasks and responsibilities rather than complete autonomous operations. However, technology-centered AI introductions that are currently still common, frequently overlook the role of the human operator and the involved organizations, resulting in significant AI effectiveness reductions and even social backlashes, slowing down progress, AI uptake, and real benefit-accrual. In the InSecTT project we have investigated various use cases for AI applications in different domains including medical, transportation, IT, and manufacturing and have formulated a framework for successful AI introductions that centers around human oversight. In this presentation we will present this framework and give examples for its application in multiple domains.
Speaker: Peter Moertl, Virtual Vehicle Research GmbH, Austria.
Dr Peter Mörtl is the manager of the human the factors teams at Virtual Vehicle Research center in Graz, Austria. His focus is on preparing and facilitating human and system integration for technological developments in automotive and AI applications, specifically focusing trust, user modeling, and user acceptance by applying human factors and systems engineering. Previously he has worked as systems engineer and human factors researcher in aviation projects in the United States to develop flight-deck and air traffic control applications. He is also Lector for Human-Systems Integration at the University of Graz, Austria.
Presentation: F01_4: Energy Efficient Deep-Edge Computing through Hardware Machine Learning.
Marcello Coppola, ST Microelectronics, France.
It is currently challenging to accomplish high data processing with low latency, because large-scale, persistent, high-bandwidth connection is often required to gather huge volumes of data from platforms and sensors operating at the edge to allow processing in the cloud. Internet-of-Things (IoT) devices deployment in remote locations for environmental monitoring with machine learning (ML) techniques, Internet-of-Medical Things, IoT-based precision agriculture, all can operate with limited battery, while robustness and reliability is required.
Speaker: Marcello Coppola, ST Microelectronics, France.
Marcello Coppola is a Technical Director at STMicroelectronics, in charge of the implementation of several research projects, including overseeing day-to-day field-based activities related to System-On-Chip Consumer Space, Automotive, IoT, Healthcare, with particular emphasis on Network-on-Chip, heterogenous multi-core hardware and software architectures which include CPU, GPU and accelerators.
17:00
End
20:00
Dinner
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session II: European Edge AI Ecosystem
9:00-9:30
Keynote presentation II_1
9:30-10:00
Keynote presentation II_2
10:00-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A02: Strategy for the Development of Edge AI Ecosystem.
Session Co-chairs:
Inessa Seifert, EPoSS, Germany, Paolo Azzoni, Inside Industry Association, the Netherlands
The goal of the session is to identify a potential roadmap for the European research and innovation in the domain of edge AI. The roadmap will coin future focus topics of the Chips Joint Undertaking and inspire other European programmes. We strive for an open ecosystem that fosters a fast adoption of research and innovation results in the European and global markets. We aim at strengthening the role of European chip vendors and system integrators in the emerging Digital Continuum and ensure high value proposition of key applications based on edge AI.
Keywords: edge AI ecosystem, European Chips Act, digital continuum, research and innovation.
Session B02: Smart Connectivity at the Edge.
Session Co-chairs:
Luca, Scuola Superiore Sant’Anna, Italy, Marcello Coppola, STMicroelectronics, France.
Edge systems will increasingly leverage collaboration to reach computational effectiveness while reducing hardware computational requirements (e.g., swarm learning). In addition, in network computation is emerging to reduce capacity requirements and closed-loop control execution time. This scenario demands for smart connectivity, both wired and wireless, at the edge where resources are allocated for data and control communications preserving energy, capacity, security and confidentiality. This session will focus on tools and techniques for providing the requested connectivity to the edge devices while guaranteeing the KPIs.
Keywords: wireless, wired, connectivity, collaborative computing, energy, capacity, security, confidentiality.
11:00-11:20
Presentation: A02_1: PREVAIL project: Test and Experimentation Facility for the Realization and Validation of Prototype Chips Dedicated to Edge AI.
Sergio Nicoletti, CEA-Leti, France.
The PREVAIL project will establish and start operating the core of a networked, multi-hub platform providing prototype chip fabrication capability, in advanced to EU stakeholders for Artificial Intelligence (AI) applications. In this project, four major European RTOs jointly work to create a multihub Test and Experimentation Facility for edge AI Hardware (TEF HW AI) build on their advanced 300 mm fabrication, design, and test facilities. Within PREVAIL, companies and research institutions will be able to develop, test and experiment with AI product prototypes based on innovative technologies developed by the four RTOs. The technologies made available are the most promising to address the specifications (power consumption, computing performance, data handling and security) required by AI chips in general and edge AI in particular.
Speaker: Sergio Nicoletti, CEA-Leti, France.
Sergio Nicoletti Ph.D joined the CEA in 2006 to develop optical gas sensors. Nicoletti has since served as chief researcher and project leader in several industrial and collaborative projects. Currently, he coordinates the TEF-Prevail Project. He earned a doctorate in physics from Université Joseph Fourier in Grenoble, France.
Presentation: B02_1: Edge AI LoRa Mesh Technologies.
Janis Judvaitis, Institute of Electronics and Computer Science, Latvia.
Intelligent connectivity at the edge combines wireless communication, edge artificial intelligence (AI), edge computing and Internet of Things (IoT) technologies to perform machine learning (ML) and deep learning (DL) tasks on connected edge devices. Low latency, ultra-low energy intelligent IoT devices with on-board computing, distributed architecture and analytics are essential as intelligent connectivity advances.
Speaker: Janis Judvaitis, Institute of Electronics and Computer Science, Latvia.
Janis Judvaitis is a researcher in Institute of Electronics and Computer Science (EDI) as well as a Ph.D student in Computer Science at University of Latvia. Works in EDI from 2013. He is involved in various international projects related to Wireless Sensor Networks, embedded systems, and Internet of Things. Expertise in ICT system design, development, implementation, and testing.
11:20-11:40
A02_2: European Ecosystem and Synergies for Fast Adoption of Edge AI Aolutions.
Presentation and discussion on the fast adoption of edge AI solutions, the development of a robust ecosystem and the creation of synergies among various stakeholders. Focus on the key elements and strategies for fostering this ecosystem and accelerating the adoption of edge AI.
Presentation: B02_2: Cascaded Look Up Table Distillation of P4 Deep Neural Network Switches.
Luca Valcarenghi, Scuola Superiore Sant’Anna, Pisa, Italy.
In-network function offloading represents a key enabler of the SDN-based data plane programmability to enhance network operation and awareness while speeding up applications and reducing the energy footprint. This paper proposes an innovative knowledge distillation technique that maps a DNN into a cascade of lookup tables (i.e., flow tables) with limited entry size. The proposed mapping avoids stateful elements and maths operators, whose requirement prevented the deployment of DNNs within hardware switches up to now. The evaluation considers a cyber security use case targeting a DDoS mitigator network function, showing negligible impact due to the lossless mapping reduction and feature quantization.
Speaker: Luca Valcarenghi, Scuola Superiore Sant’Anna, Pisa, Italy.
Luca Valcarenghi is an Associate Professor at the Scuola Superiore Sant’Anna of Pisa, Italy, since 2014. He received the Laurea in Electrical Engineering in 1997 from Politecnico di Torino and the M.S.E.E. and Ph.D. in Electrical Engineering Major Telecommunications from UTD in 1999 and 2001, respectively. He published more than three hundred papers in International Journals and Conference Proceedings. Dr. Valcarenghi received a Fulbright Research Scholar Fellowship in 2009 and a JSPS “Invitation Fellowship Program for Research in Japan (Long Term)” in 2013. His main research interests are optical networks design, analysis, and optimization; communication networks reliability; energy efficiency in communications networks; optical access networks; zero touch network and service management; 5G technologies and beyond.
11:40-12:00
A02_3: Collaboration Along the AI Value Chain
Collaboration along the AI value chain is crucial for the development, deployment, and success of AI solutions. The discussion will focus on which are the AI value chain various stages, from data collection and preprocessing to model development, deployment, ongoing optimization and what are the catalysts for increasing the collaboration.
Presentation: B02_3: Energy Optimisation for Edge Devices.
Javad Chamanara, University of Hannover, Germany.
The presentation gives an overview of RLKube, an RL-based custom Kubernetes scheduler plugin for optimized task scheduling. RLKube’s objective is to improving energy efficiency in a Kubernetes cluster. RLKube considers multiple optimization objectives, including energy optimization and Pod throughput. It utilizes DDQN with PER with different reward functions to train the RL agent. Our evaluations indicate the effectiveness of RLKube in outperforming the default Kubernetes scheduling policies in terms of throughput and energy usage.
Speaker: Javad Chamanara, University of Hannover, Germany.
Dr. Eng. Javad Chamanara has a PhD computer science. He is a research team lead at the University of Hannover Germany. His research projects are focused on federated learning, energy optimization, and privacy-aware computing. He is a participating to the CLEVER, BRAINE, and Glaciation projects to provide privacy-aware, energy optimization to federated data operations.
12:00-12:20
A02_4: Market Impact Along AI Value Chain for Co-creation of Edge Applications.
The discussions will focus on the co-creation of edge applications, where multiple stakeholders collaborate along the AI value chain to develop and deploy edge AI solutions, and how this co-creation have significant market impacts.
Presentation: B02_4: Edge-assisted Smart Agriculture for High Yield Eco-farms.
Alessandro Pacini, Scuola Superiore Sant’Anna, Italy.
This presentation describes the architecture and lists a preliminary set of requirements of an edge-assisted smart agriculture application for high yield eco-farms. In particular, requirements in terms of computational resources, energy, latency, transmission capacity are considered in the specific deployment of the CLEVER project.
Speaker: Alessandro Pacini, Scuola Superiore Sant’Anna, Italy.
Alessandro Pacini is a PhD student in Emerging Digital Technologies at the Scuola Superiore Sant’Anna. He received his bachelor’s degree in computer science from the University of Camerino in 2018. He then pursued his studies and obtained a joint master’s degree in computer science and Networking at the University of Pisa and Scuola Superiore Sant’Anna in 2021. During this time, he won a one-year research fellowship at SSSA focused on building a scalable and reliable monitoring architecture for optical networks. His main research interests are in the field of next-generation SD networks, with a particular focus on reusing existing network architectures to move towards a zero-touch paradigm.
12:30-13:30
Networking Lunch
13:30-15:30
Session C02: Edge AI Technology Development.
Session Co-chairs:
Ovidiu Vermesan, SINTEF AS, Norway, Patrick Pype, NXP Semiconductors, Belgium.
Edge AI combines IoT, edge computing and AI that implements new concepts for state-of-the-art AI applications at the edge. The edge technologies push the frontiers of AI to the network edge to fully unleash the potential of real-time data processing. The goal of the session is to present the latest trends and challenges.
Keywords: edge AI technologies, edge AI granularity, edge AI continuum, edge AI technology stack.
Session D02: Agri-Food and Beverage.
Session Co-chairs:
Markus Tauber, Research Studios Austria, Austria, Aymeric Deshayes, Laboratoire IMS, France
AI at the edge, on devices located closer to the data source, such as sensors and machinery within the agricultural and food processing sectors is advancing rapidly. It brings intelligence and automation to various processes, optimising efficiency, reducing waste, improving quality, and ensuring sustainability. The session covers several innovative edge AI applications in the domain with the goal of improving sustainability, reducing waste, enhancing quality, and ensuring agr-food and beverage safety throughout the entire supply chain.
Keywords: precision agriculture, energy efficiency, safety and traceability, resource optimisation, real-time decision support, yield prediction, disease identification, crop production optimisation.
13:30-13:50
C02_1 – Edge AI Frameworks Advancements.
Presentation and discussions on the trends on the development of edge AI frameworks and the developments to support edge deployments.
Presentation: D02_1: Using Edge AI in IoT Devices for Smart Agriculture Applications: Autonomous Weeding of Vegetables and Vine Health Monitoring.
Aymeric Deshayes, Laboratoire IMS, France.
Crop production uses increasingly fixed sensors in the plots, on-board sensors on tractors and robotic machines. These devices frequently have limited access to networks which make EDGE architectures particularly suitable. Through two real life use cases, this article shows how applications based on artificial vision can benefit from electronic circuits, making it possible to use artificial intelligence in low-cost EDGE architectures.
Speaker: Aymeric Deshayes, Laboratoire IMS, France.
Aymeric Deshaye is a research engineer who graduated in 2017 from the agronomic school of Bordeaux Sciences Agro, with a specialization in digital for agriculture. He is currently working at the IMS research Lab. (Bordeaux – France) in a group working more specifically on image processing applied to agriculture. Aymeric Deshayes is involved in various research projects including vine disease detection, robotic weeding and Edge Computing for Crop Monitoring. He is currently project manager for IMS Lab within the ANDANTE Project.
13:50-14:10
C02_2 – Immersive Technologies and Edge AI Convergence.
Presentation and discussions addressing the convergence of immersive technologies and edge AI has the potential to create transformative experiences and applications across various industries. Natural and responsive interaction within immersive environments. Gesture recognition, voice commands, and facial expression analysis can be processed locally, allowing for quicker responses and a more intuitive user experience.
Presentation: D02_2: Vineyard Yield Prediction.
Lilian Hollard, Université de Reims Champagne-Ardenne, France.
The Value Chain Agri-food and Beverage initiative strives to equip the European Agri-food and Beverage industry with pioneering technologies, protocols, and AI algorithms. This endeavour seeks to establish the industry’s global eminence through its leadership, resilience, and contributions to achieving climate neutrality. Within the framework of this project, we showcase the latest advancements in technology and applications, with a particular emphasis on AI and edge AI methodologies. These methodologies and techniques, used predominantly in the context of vineyard yield forecasting, stand out in the yield forecasting demonstrator, which is dedicated to crafting a dependable edge AI framework for precise yield forecasting, striving for an error margin of less than 15%.
Speaker: Lilian Hollard, Université de Reims Champagne-Ardenne, France.
First year Ph.D. student at LICIIS from the University of Reims Champagne Ardenne, development of Artificial Intelligence models for viticulture: Disease identification and yield estimation. Adaptation and optimization of Deep Learning models for Edge and mobile devices.
14:10-14:30
C02_3 – Edge AI in Autonomous Systems.
The discussions will address the edge AI technology developments for autonomous systems and robotics including specific issues related to object recognition, navigation, and human-robot interaction.
Presentation: D02_3: Disease Identification for Vineyards Fields.
Lucas Mohimont, Université de Reims Champagne-Ardenne, France.
Disease Identification aims to progress with an advanced edge AI framework tailored for real-time disease detection in the field. The latest techniques are implemented within an object detection architecture, effectively enhancing the deep learning model for tomato disease prediction. This enhancement enables real-time, on-site predictions, facilitating efficient large-scale processing. In the initial stages, meta-edge analysis could contribute to generating a map of affected zones, offering champagne manufacturers and winegrowers the means to optimize the application of phytosanitary products. The presentation focuses on state-of-the-art technological and application developments with a spotlight on AI and edge AI methods and techniques used in the project.
Speaker: Lucas Mohimont, Université de Reims Champagne-Ardenne, France.
Ph.D. in Artificial intelligence for industry. Application to viticulture and COVID-19 case modelling. Lucas Mohimont is working at the University of Reims Champagne Ardenne for the development of Artificial Intelligence models for viticulture: Disease identification and yield estimation.
14:30-14:50
C02_4 – Edge AI Standardisation Activities.
Erwin Schoitsch, AIT Austrian Institute of Technology, Austria.
Presentation and discussions on standardisation activities, EU AI Act and the challenges related to verification, validation and testing of edge AI systems.
Standardization is key to a fair, competitive society and economy. This includes technical standards, specifications and reports developed by international, European and/or industrial alliances in generic and domain-specific areas of interest. Of public interest and for public acceptance of new technologies are issues concerning safety, dependability, cybersecurity and ethical concerns, particularly in context of decision making algorithms and Artificial Intelligence. On European level, the carbon-footprint and “Green Deal” contributions play an important role as well. The presentation will provide an overview on the current efforts in standardization and by publishing guidelines and recommendations, and how they are addressed in the project AI4CSM (Automotive Intelligence for Connected Shared Mobility).
Erwin Schoitsch works more than 50 years in different positions for AIT Austrian Institute of Technology. Since his retirement in 2010 he works as Free Lancer under various subcontracts in many European Research projects for AIT and other organizations. Focus of his engagement was always on trustworthiness of highly reliable, safety-critical systems (Dependability, Safety, Cybersecurity), in industrial projects as well as European Research. He is active in many standardization groups of ISO, IEC and ISO/IEC JTC1, and the Austrian National Mirror Committees, covering topics like functional safety, cybersecurity, automated driving, Artificial Intelligence and ethical concerns. He is in several European Research Projects leading the standardization tasks.
Presentation: D02_4: CamFridge: Edge vs. Cloud Implementation of Object Recognition in the Fridge with Camera Use Case.
Ebrar Ömer, Arcelik, Turkey.
This study addresses a pivotal choice: adopting edge (local devices) or cloud computing for our smart camera-equipped fridge. Cloud systems utilize remote servers, while edge computing processes data instantly on devices. We aim to assess the YOLOv7 model’s efficacy in both paradigms for our smart fridge, given their distinct merits and challenges.
Speaker: Ebrar Ömer, Arcelik, Turkey.
Ebrar Ömer, is a 23-year-old computer engineering graduate from Istanbul Technical University (İTÜ). Currently, she works at Arçelik and has a deep passion for artificial intelligence and machine learning. Fluent in both Turkish and English, she aspires to build her career in these dynamic fields, contributing to the ongoing transformation of the technology landscape.
15:00-15:30
Afetrnoon Coffee/Tea Break
15:30-17:00
Session E02: Energy Optimisation for Edge Devices.
Session Co-chairs:
Pavel Smrz, COGNITECHNA s.r.o., Czech Republic, Paolo Meloni, Università degli studi di Cagliari, Italy
This session is dedicated to the research on power-aware edge computing, which aims to optimize edge devices’ energy efficiency and performance by dynamically adapting to the workload, communication needs, and environmental conditions. Presented papers will discuss the benefits of running optimised AI-based solutions regarding decreased power requirements and long-life operation for near real-time data processing.
Keywords: power awareness, energy efficiency, micro-power management, long battery life, dynamic adaptation to workload, energy-optimised AI, machine learning for energy optimisation.
Session F02: Edge AI Hardware Platforms.
Session Co-chairs:
Björn Debaillie, imec, Belgium, Olaf Peters, Technolution, the Netherlands.
Edge AI hardware platforms are specialised computing systems designed to accelerate and optimise the execution of AI workloads on edge devices. Various platforms are used for running AI models efficiently at the edge, where resource constraints, power efficiency, and real-time processing requirements. The session gives an overview of different technological developments addressing several edge AI HW platforms.
Keywords: multi-application HW platforms, FPGAs, Soc, SoM, MCUs, heterogeneous integration, SNN processor platforms, automated optical inspection edge AI inference platforms, heterogeneous edge systems, platforms for federated learning.
15:30-15:50
Presentation: E02_1: Tiny Transformers for On-line Processing of Sensor Data Streams on Low-power MCUs.
Paola Busia, Università degli Studi di Cagliari, Italy.
The presentation discusses the design, training and deployment of lightweight energy efficient algorithms derived from vision transformers exploited for the real-time analysis of sensor data streams. The capabilities of the presented approach are assessed on decoding and classification tasks involving biosignals. The algorithms are executed on several parallel MCU for IoT-node development.
Speaker: Paola Busia, Università degli Studi di Cagliari, Italy.
Paola Busia is assistant professor at Università degli studi di Cagliari. Her main research interests involve power efficient processing architectures for embedded applications and for at-the-edge AI.
Presentation: F02_1: SNNs on FPGAs for On-line Processing of Sensor Data.
Paolo Meloni, Università degli Studi di Cagliari, Italy.
The presentation discusses the design, training and deployment of lightweight energy efficient spiking neural networks exploited for the real-time analysis of sensor data streams. The capabilities of the presented approach are assessed on decoding and classification tasks involving biosignals. The algorithms are executed on a custom SNN processor deployed on low/mid-end FPGAs.
Speaker: Paolo Meloni, Università degli Studi di Cagliari, Italy.
Paolo Meloni is associate professor at Università degli studi di Cagliari. His main research interests involve power efficient processing architectures for embedded applications and for at-the-edge AI.
15:50-16:10
Presentation: E02_2: CA-CFAR is Convolution: Fast Target Detection With Machine Learning Accelerator.
Chen Liu, Technische Universität Dresden, Germany.
In radar target detection, Constant False Alarm Rate (CFAR) is a commonly employed detector known for its simplicity and effectiveness. Its sliding-window detection mechanism possesses computational similarity to convolutional operations in machine learning. With the increasing emergence of AI-enhanced radar processing algorithms, systems at the edge are often equipped with ML accelerators to expedite matrix multiplications and convolutions. This paper introduces a heuristic algorithm that equivalently maps Cell-Averaging (CA) CFAR to a convolutional operation and leverages machine learning accelerators for acceleration. Comparative experiments against the conventional CA-CFAR algorithm executed on ARM cores demonstrate that our proposed method significantly reduces the processing latency by 38 times without incurring substantial memory overhead. This advancement brings substantial promise for potentiating real-time high-resolution radar target detection.
Speaker: Chen Liu, Technische Universität Dresden, Germany.
Chen Liu received the BSc degree and MSc degree in electrical engineering in 2012 and 2016, respectively. He is currently working toward the PhD degree in Technische Universität Dresden, Germany. His research interests include intelligent radar signal processing, efficient deep learning on the edge and neuromorphic computing in automotive.
Presentation: F02_2: ANDANTE Platform 4.1a : A Multi-Application Platform Supporting Several Uses Cases in the Domains Digital Farming and Transport and Smart Mobility.
Ivan Miro-Panades, CEA, France.
An IoT challenge is to provide devices supporting various applications to reduce cost and parts management. As part of ANDANTE, a multi-application platform has been developed to address 6 use cases in two different applications domains Digital Farming and Transport and Smart Mobility. Platform 4.1a is a combination of an AI ASIC accelerator, NeuroCorgi, and a commercial FPGA including an ARM core. The implementation of this platform required to make a compromise between the FPGA and the ASIC to share the ANN layers and to optimise them according to the needs of the application to guarantee high application performance (energy efficiency, latency, accuracy, and good level of integration).
Speaker: Ivan Miro-Panades, CEA, France.
Ivan MIRO-PANADES received the M.S. degree in telecommunication engineering from the Technical University of Catalonia (UPC), Barcelona, Spain, in 2002, and the M.S. and Ph.D. degrees in computer science from Pierre and Marie Curie University (UPMC), Paris, France, in 2004 and 2008, respectively. He worked at Philips Research, Sureness, France and STMicroelectronics, Crolles, France, before joining CEA, Grenoble, France, in 2008, where he is currently an Expert Research Engineer in digital integrated circuits. His main research interests are artificial intelligence, the Internet of Things, low-power architectures, energy-efficient systems, and Fmax/Vmin tracking methodologies.
16:10-16:30
Presentation: E02_3: Radar-based Reactive Planning for Drones.
Hamed Javadi, imec, Belgium.
While cameras and lidars are more common in reactive planning in drones, radars can potentially complement, or replace, them thanks to their robustness against harsh weather and extreme light conditions. To this end, appropriate signal processing algorithms are needed to tackle the poor angular resolution of radars. These algorithms, on the other hand, should fit the limitations at edge (drone). In this presentation, we present a computationally efficient synthetic aperture radar (SAR) pipeline as a solution to radar-based reactive planning in drones.
Speaker:Hamed Javadi, imec, Belgium.
S. Hamed Javadi is working as a researcher at imec, Belgium, in the fields of signal processing, sensor fusion, and AI. His current work focuses on the development of AI algorithms for radar signals which also requires strong radar signal processing as a preprocessing step. Before joining imec, S. Hamed Javadi was a postdoc at Ghent University, Ghent, Belgium, where he developed efficient AI-based algorithms for smart farming based on multiple sensors’ data. S. Hamed Javadi was also an assistant professor at the University of Bojnord, Iran, where his main focuses of interest were statistical signal processing, distributed algorithms, and machine learning.
Presentation: F02_3: Automated Optical Inspection Edge AI Inference System Solutions.
Claudio Cantone, High Technology Systems H.T.S. srl, Italy and Alberto Faro, Deepsensing, DEEPS, Italy.
The presentation gives an overview of the automated optical inspection edge AI inference system solutions in digital industry by discussing if and how they allow the manufacturers to achieve a satisfactory trade-off between customer requirements and production costs. The comparison of the available solutions using KPIs, FRs and NFRs points out that AOI consisting of testing boards at the edge based on a model pre-trained on the cloud server is a promising solution if it is provided with a suitable worker interface and planning tools for managers.
Speaker: Claudio Cantone, High Technology Systems H.T.S. srl, Italy and Alberto Faro, Deepsensing, DEEPS, Italy.
Claudio Cantone received the Electronic Engineering degree from University of Pisa and the Master in Business Administration from Bocconi University. Initially he worked for several years at Philips Spa, Cinisello Balsamo, Milan in the sector of Electromedical Technical Services. Then he founded HTS srl that is a company providing Electronic Manufacturing Services (EMS) to support the entire life cycle of products in digital industry. Currently he is the Head of the HTS R&D Department.
Alberto Faro received an engineering degree from Milan Polytechnic. He was one the scientists who designed and tested the first European Informatics Network (EIN), forerunner of the Internet. From 1980 to 2016 he was Professor of Artificial Intelligence at the University of Catania where he contributed to find the Electronic and Computer Engineering Degrees. In 2019 he founded Deepsensing srl that is an innovative Startup active in the Artificial Intelligence of Things (AIoT) field.
16:30-16:50
Presentation: E02_4: Inside the AI Accelerators: From High Performance to Energy Efficiency.
Ana Pinzari, Univ. Grenoble Alpes, France.
This presentation covers current technologies for high-performance, low-power neural networks. To cope with the high computational and storage resources, hardware optimisation techniques are proposed: DL compilers and frameworks, DL hardware and hardware-specific code generator (FPGA). More specifically, in this presentation we explore the quantization mechanism in deep learning, based on a deep-CNN classification model. We highlight the accuracy of quantized models and explore their efficiency on a variety of hardware platforms. Through experiments, we show the performance achieved using general-purpose hardware (CPU and GPU) and a custom ASIC (TPU), as well as the simulated performance for a reduced bit-width representation of 4 bits, 2 bits (ternary) up to 1-bit heterogeneous quantization (FPGA).
Speaker: Ana Pinzari, Univ. Grenoble Alpes, France.
Ana Pinzari received her Ph.D degree in computer science at Université de Technologie de Compiègne, Compiègne, France in 2012, in Real-Time Embedded Systems. Then, she worked in a scientific research organization ECSI (European Electronic Chips & Systems Initiative) focusing on new design methods, tools, and standards for design of complex electronic systems. In 2021, she joined Grenoble INP, TIMA, as a post-doctoral research engineer. Her current research focuses on exploring optimisation methods for the hardware implementation of neural networks.
Presentation: F02_4: Federated Learning in Cybersecurity.
Dimitrios Serpanos, CTI and University of Patras, Greece.
Federated learning constitutes a promising solution for ML/AI-enabled cybersecurity solutions in distributed environments with Edge systems. We present an efficient malware detection solution employing federated learning for environments that combine powerful as well as limited resource edge systems. We demonstrate that federated learning enables effective cybersecurity solutions for distributed environments that include heterogeneous edge systems and have constraints and limitations in data exchange, ranging from performance to IP.
Speaker: Dimitrios Serpanos, CTI and University of Patras, Greece.
Dimitrios Serpanos is president of the Computer Technology Institute & Press DIOPHANTUS and Professor of ECE at the University of Patras, Greece. He is Chair of the Scientific Council of INSIDE Industrial Association.
17:00
End
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session III: Edge AI Applications
9:00-9:30
Keynote presentation III_1
9:30-10:00
Keynote presentation III_2
10:00-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A03: Intelligent Mobility Systems.
Session Co-chairs:
Christian Ohms, Mercedes-Benz AG, Germany, Alfred Höß, OTH Amberg-Weiden, Germany.
The integration of AI capabilities at the edge of a network or within vehicles, transportation, and mobility systems to enable autonomous mobility to enhance the capabilities and increase safety and reliability. These systems leverage AI algorithms and real-time data processing to make decisions, control vehicles, and navigate autonomously without relying heavily on remote cloud-based computation. The session includes presentations that address key aspects and applications of autonomous edge AI mobility systems.
Keywords: autonomous vehicles, edge AI for perception, path planning, localisation, advanced driver assistance systems, V2X communication.
Session B03: Digital Industry.
Session Co-chairs:
Falk Hoffmann, NXP Semiconductors, Germany, Claus Lenz, Cognition Factory GmbH, Germany.
The transformation and integration of digital technologies, such as data analytics, automation, the IoT, AI, and immersive technologies require the deployment of these technologies into various sectors of the economy and industrial processes. The session presents various use case employing edge AI technologies in industrial manufacturing processes that drive technological advancements and changing manufacturing patterns with a profound impact on the economy, markets, and society.
Keywords: advanced AI-based manufacturing, maintenance, repair, and operations, functional safety edge AI systems, intelligent radar signal processing.
11:00-11:20
Presentation: A03_1: AI for Connected and Shared Mobility.
Jochen Koszescha, Infineon Technologies AG, Munich, Germany.
The presentation is giving an overview of AI4CSM project that focuses on the design of highly optimized and automated components for digital mobility that allows smart and energy efficient fleet operations in terms of smart routing and platooning achieving low latency in the vehicle intelligence and perception systems. This means not only smart sensor fusion, but also reducing the level of complexity by using functional architectures and trustable automotive AI, to foster Europe’s competitiveness in the global automotive market. The vision of AI4CSM is to build Europe’s intelligent electronic component and systems for ECAS 2030 vehicles supporting European mass market production, manufacturability and scalability based on the Green Deal principles (incl. Vision Zero and Safe System) to address the sustainable urbanization challenge defined by the United Nations.
Speaker: Jochen Koszescha, Infineon Technologies AG, Munich, Germany.
Jochen Koszescha is a senior director at Infineon Technologies AG with over 30 years of experience in Electronic Components and Systems at the Aviation and Transport Industries. He has been coordinating several research and development projects at Infineon.
Presentation: B03_1: Distributed Multisensory System, from Pre-Development to Production.
Claus Lenz, Cognition Factory GmbH, Germany.
This talk showcases the industrial use-case of a distributed multisensory system applied in the field of Maintenance, Repair, and Operations. It highlights the advantages and challenges of a distributed architecture of sensor nodes, capable of independent functioning and data processing. The presentation covers the development cycle from the Requirement formulation phase, up to Deployment and Maintenance. Emphasis is made on certifiable approach to AI-based solutions, and delegation of intelligence to the edge, as well as an architectural design that is scalable both vertically and horizontally.
Speaker: Claus Lenz, Cognition Factory GmbH, Germany.
CEO and co-founder of Cognition Factory GmbH. PhD in human-robot collaboration at Technische Universität München (TUM). Experienced within research projects aimed at digitising the industry: AI4DI, Providentia++, DiProLea.
11:20-11:40
Presentation: A03_2: Natural Language Conditioned Planning of Complex Robotics Tasks.
Janis Arents, Institute of Electronics and Computer Science, Riga, Latvia.
As natural language processing advances in the field of robotics, enabling seamless human-robot interaction, it becomes imperative to identify the most effective approach for conditioning complex robotics tasks using natural language commands. This chapter reviews various state-of-the-art methods for natural language-conditioned planning, with a particular focus on mobile manipulation. We explore and review different architectures and techniques to comprehend, interpret, and execute natural language commands. Challenges are identified along the way, and conceptual architecture is proposed to tackle them in an efficient manner.
Speaker: Janis Arents, Institute of Electronics and Computer Science, Riga, Latvia.
Researcher Janis Arents has been working in EDI since 2016. Graduated Riga Technical University (Riga, Latvia) in the year 2018 with professional master’s degree in electrical engineering. Currently studying in Riga Technical University in Computerised Control of Electrical Technologies doctors’ degree program. Main research topics are connected with innovative digital technology usage in industrial applications, including computer vision and artificial intelligence-based systems for industrial process automation.
Presentation: B03_2: Outlook on the Deployment and Execution of Mixed-Criticality Edge-AI Applications.
Suzana Veljanovska, Zurich University of Applied Sciences, Switzerland.
Functional safety-relevant edge AI systems are the cutting of AI research, and the technological gap is significant. In this presentation we approach the issue of creation and deployment of both typical AI applications and functional safety systems. We pay particular attention to the merging of the two workflows and present options and ideas to achieve this.
Speaker: Suzana Veljanovska, Zurich University of Applied Sciences, Switzerland.
Suzana Veljanovska is a research assistant at Zurich University of Applied Sciences (ZHAW) working in the field of dependable AI.
11:40-12:00
Presentation: A03_3: Dynamic LoRa Broadcasting and Obstacle Prediction.
Dimitris Georgiadis, Harokopio University of Athens, Greece.
This presentation covers significant achievements in the development of a LoRa-based emergency communication network for highly automated vehicles, validated through simulations using NS3. The simulation comprises stationary infrastructure, including a gateway a network server, and multiple mobile vehicles acting as end devices. Concluding in this research project we present a significant advancement in LoRa-based emergency communication simulations, with achievements in realistic mobility modelling, communication protocols, and visualization tools. Ongoing efforts are focused on enhancing emergency responsiveness focused on obstacle recognition, optimizing communication reliability, and accelerating the network’s adaptability to dynamic scenarios.
Speaker:Dimitris Georgiadis, Harokopio University of Athens, Greece.
Dimitris Georgiadis is a research associate at Harokopio University of Athens. His works focus on 5G D2D single-cell communications and minimization of energy consumption, UAV swarm communications and currently studying vehicle behavior. Graduate from Harokopio University of Athens from the Department of Informatics and Telematics and is now a postgraduate Student of Master Programme (MSc) MPhil in Computer Science and Informatics.
Presentation: B03_3: RANet: An Autoencoder for High-Resolution Range-Angle Maps in Multistatic Radar.
Chen Liu, Technische Universität Dresden, Germany.
Advanced radar perception tasks, such as target recognition and tracking, often impose strict requirements on radar imaging quality. However, conventional radar imaging methods are constrained by spatial angular resolution, which is limited by the number of virtual channels composed by the antenna array, thus leading to suboptimal performance of low-cost, low-resolution radar systems. This paper introduces an innovative deep learning model: RANet, which leverages the raw sampling signals from multiple radar sensors to generate high-resolution range-angle maps in an end-to-end manner. RANet not only partially replaces the traditional radar processing chain but also demonstrates superior performance in downstream tasks such as object counting and tracking. Experimental results indicate that monitoring up to four indoor individuals achieves an accuracy of 96.4% by exploiting accumulated temporal information of only two 1T3R (one transmitter, three receivers) radar sensors.
Speaker: Chen Liu, Technische Universität Dresden, Germany.
Chen Liu received the BSc degree and MSc degree in electrical engineering in 2012 and 2016, respectively. He is currently working toward the PhD degree in Technische Universität Dresden, Germany. His research interests include intelligent radar signal processing, efficient deep learning on the edge and neuromorphic computing in automotive.
12:00-12:20
Presentation: A03_4: Edge Perception in Autonomous Driving.
Gerhard Dorn, Virtual Vehicle Research GmbH, Austria.
The talk gives a small introduction to our current work on the autonomous driving Robotaxi within the EU project AI4CSM and focusses on development and integration techniques of edge perception algorithms using Julia for machine learning tasks. Challenges regarding data quality, data sources, uncertainties, and run-time issues as well as opportunities regarding fast prototyping and integration will be discussed.
Speaker:Gerhard Dorn, Virtual Vehicle Research GmbH, Austria.
Gerhard Dorn is a senior scientist at Virtual Vehicle Research GmbH with a background in mathematics and computational physics.Within the project AI4CSM Gerhard works on LIDAR perception and the integration into the test vehicle Ford Mondeo.
Presentation: B03_4: Smart Semiconductor Manufacturing Assistant (SSMA) – An Efficient Quality Management Supporting Tool.
Silke Braun, Infineon, Austria.
Smart Semiconductor Manufacturing Assistant (SSMA) as an efficient quality management support tool is an important step in the direction of establishing causal links between production-related data such as production equipment and process parameters, data resulting from process experiments as well as knowledge from FMEAs. This allows potential risks to be quickly identified, predicted, and classified to support the engineer in decision making.
Speaker: Silke Braun, Infineon, Austria.
Silke Braun completed her studies in microsystems engineering at Mittweida University of Applied Sciences (Germany). In 2017, she started working at Infineon Technologies Dresden as a defect density and metrology engineer; since 2018 in Villach (Austria). She strengthened her knowledge in the field of quality management, among others for new technologies like GaN. Since 2022, she focuses her work on the field of funding and works in various EU-funded projects.
12:30-13:30
Networking Lunch
13:30-15:00
Session C03: Autonomous Vehicle Domains Applications.
Session Co-chairs:
Jochen Koszescha, Infineon Technologies AG, Germany, George Dimitrakopoulos, Harokopio University of Athens, Greece.
The integration of AI capabilities at the edge of a network or within vehicles, transportation, and mobility systems to enable autonomous mobility to enhance the capabilities and increase safety and reliability. These systems leverage AI algorithms and real-time data processing to make decisions, control vehicles, and navigate autonomously without relying heavily on remote cloud-based computation. The session includes presentations that address key aspects and applications of autonomous edge AI mobility systems.
Keywords: autonomous vehicles, edge AI for perception, path planning, localisation, advanced driver assistance systems, V2X communication.
Session D03: Digital Society.
Session Co-chairs:
Fetze Pijlman, Signify, the Netherlands, Silke Braun, Infineon, Austria.
The deployment of AI capabilities directly at the edge of the network or on devices within a digitally connected society plays a pivotal role in shaping how technology interacts with individuals and communities, enabling real-time decision-making, enhancing user experiences, and ensuring data privacy. The session focus on several presentations that highlight the use of edge AI in digitising the society by promoting efficiency, responsiveness, and data privacy while enhancing the quality of life and safety for individuals and communities advancing the development of a more connected and intelligent world.
Keywords: smart working environments, digital healthcare, augmented reality, event-based intelligence, environmental monitoring.
13:30-13:50
Presentation: C03_1: SNN-based Accelerators for Low-latency, Ultra-low-power Inference at the Edge.
Maxim Rykunov, imec, Belgium.
State-of-the-art currently focuses on conventional deep neural network architectures for classification lacking in power, performance, and scalability especially for the processing at the edge. In this talk we present two designed and evaluated neuromorphic prototypes to be demonstrated in the AI4CSM project. One of these solutions focuses on a HW/SW co-design based on a spiking neural network and embedded processing for automatic detection, tracking and classification of multiple vulnerable road users. The second one targets the development of a programmable analogue-mixed signal accelerator platform for always-on detection and recognition of time-series patterns in sensor data. Neuromorphic SNN-based approaches are critical in such applications, where we demonstrate novel energy-efficient platforms based for low-latency, ultra-low-power inference at the edge.
Speaker: Maxim Rykunov, imec, Belgium.
Maxim Rykunov received the Ph.D. degree in electrical and electronic engineering from Newcastle University, Newcastle, U.K., in 2013, in asynchronous microprocessors and design of power efficient systems. In 2014, he joined IMEC, Leuven, Belgium, where he is currently a Project Manager in Advanced RF group. His research interests include real-time embedded systems for heterogeneous multisensor data fusion and energy-efficient machine learning for high-resolution radars.
Presentation: D03_1: On the Move Biometrics at the Edge: Security Meets Convenience.
Emanuele Maiorana, Roma Tre University, Italy.
Biometric recognition systems are used to validate the identity of an individual and among the main candidates to be used in contexts with high security requirements (e.g., airports, or border control). Interacting with a biometric recognition system requires a user to stop at a terminal for the time necessary to acquire the relevant traits. In contexts requiring the management of large flows of people, such procedure could lead to the creation of long lines, resulting in discomfort for the involved subjects. To deal with this issue, we propose an “on-the-move” solution where subjects to not need to stop at a gate during the recognition process, since the interested characteristics can be captured while users continue to walk, thus increasing the system throughput and improving passengers’ convenience. The required processing is performed at the edge, thus realising a reliable, low-cost, and convenient system to perform security checks within an airport. In more detail, the proposed system relies on a multimodal approach exploiting face traits to trigger the recognition process at a distance, and subcutaneous vein patterns to guarantee the desired robustness against malicious spoofing attempts.
Speaker: Emanuele Maiorana, Roma Tre University, Italy.
Emanuele Maiorana received the Ph.D. degree in biomedical, electromagnetism, and telecommunication engineering from Roma Tre University, Rome, Italy, in 2009, with European Doctorate Label. He is currently an Assistant Professor with the Department of Industrial, Electronic and Mechanical Engineering, Roma Tre University. His research interests are in digital signal and image processing, with a specific emphasis on biometric recognition. Dr. Maiorana was a recipient of the Lockheed Martin Best Paper Award for the Poster Track at the IEEE Biometric Symposium 2007, the Honeywell Student Best Paper Award at the IEEE Biometrics: Theory, Applications and Systems Conference 2008, and the Best Paper Award at the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM) 2022. He was the General Chair of the 9th IEEE International Workshop on Biometrics and Forensics (IWBF) 2021. He is an Associate Editor of the IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY.
13:50-14:10
Presentation: C03_2: Robust Propulsion System for Shared Connected Mobility.
Christian Ohms, Mercedes-Benz AG, Germany.
The talk will address developments in drivetrain technology with regard to efficiency, diagnostic methods and charging components as well as current trends in the automotive industry.
Speaker: Christian Ohms, Mercedes-Benz AG, Germany.
Christian Ohms received his diploma in electrical engineering from the University of Applied Sciences in Kiel in 2008. After holding different engineering positions in the automotive industry, he joined the group of power electronics advanced engineering at Daimler AG, now Mercedes-Benz AG, in 2015. His main tasks include component design for future drivetrain concepts and coordination of vehicle demonstrator projects. He has been actively collaborating for various funded projects like HV-ModAL, 3CCar and AutoDrive.
Presentation: D03_2: AI-enabled Resource Optimization in Critical Computing Infrastructures.
Emmanouel (Manos) Varvarigos, National Technical University of Athens, Greece.
The presentation will give an overview of the AI-enabled resource optimisation for critical computing infrastructures, enabling to improve the IT infrastructure and enabling greater productivity, efficiency, streamline operations and reduce redundancy.
Speaker: Emmanouel (Manos) Varvarigos, National Technical University of Athens, Greece.
Emmanouel (Manos) Varvarigos received a Diploma in Electrical and Computer Engineering from the National Technical University of Athens (NTUA) in 1988, and the M.S. and Ph.D. degrees in Electrical Engineering and Computer Science from the Massachusetts Institute of Technology (MIT) in 1990 and 1992, respectively. He has held faculty positions at the University of California, Santa Barbara and Delft University of Technology, the Netherlands. He was a Professor of Computer Engineering and Informatics (CEID) at the University of Patras. In June 2015 he was elected Professor of Electrical and Computer Engineering at the National Technical University of Athens, Greece, where he created the High-Speed Communication Networks Lab.
14:10-14:30
Presentation: C03_3: Reward Shaping for Safe Reinforcement Learning Controllers.
Lorenz Klampfl, AVL List GmbH, Austria.
The performance of controllers based on reinforcement learning heavily depends on designing appropriate reward functions. With poorly designed rewards, an agent can behave unsafely, non-optimal, learn how to obtain the reward without achieving the desired goals, or take too long to learn a policy. In this talk, we give an overview of formal specifications to automatically provide locally shaped rewards. The advantage of such an approach is that we can address all the above-mentioned challenges efficiently.
Speaker: Lorenz Klampfl, AVL List GmbH, Austria.
Lorenz Klampfl holds an M.Sc. in Aviation (2017), is currently pursuing a Ph.D. in Computer Sciences at Graz University of Technology and works as a Simulation Engineer at AVL List GmbH. He is conducting research on quality assurance methodologies for autonomous cyber-physical systems, including the reliability and trustworthiness of neural networks and reasoning techniques for autonomous systems, at the Institute of Software Technology. In his role as a Simulation Engineer, he focuses on the development of ADAS/AD verification and validation methodologies in virtual environments as well as on the edge.
Presentation: D03_3: Quantization- and HW-aware Neural Architecture Search for EdgeAI Solutions.
Gerardo Daalderop, NXP Semiconductors, the Netherlands.
To a significant extent, future smartness in devices of the digital society will originate from neural networks embedded in these devices. However, today typically neural network (NN) models are solely trained to improve the performance of the prediction-the primary task-, without considering the embedded constraints of the devices in which these models will be deployed. Afterwards in a separate process, the models are adapted to fit memory, computation, and performance requirements. The presentation introduces the concept of multi-objective quantization-aware neural architecture search, wherein optimization techniques like quantization and pruning, are combined with HW-aware neural architecture search (NAS). NAS has become a very successful technique for optimizing neural networks, and in HW-aware NAS the properties of the embedded device are brought into the loss-function and training loop, allowing optimization both towards the primary task, as well as secondary objectives being non-functional requirements. A few examples will be given highlighting the benefits of the approach.
Speaker: Gerardo Daalderop, NXP Semiconductors, the Netherlands.
Gerardo Daalderop (PhD, MBA) is a technical director at NXP Semiconductors, having in recent years built-up an AI ML Innovation Center in its Central Technology Office, focusing on AI ML innovations for the Edge beyond next generation products, and across segments as mobility, industrial, home/office, IoT. Before this, he has set up and led groups, roadmaps, programs, and projects in the areas of highly automated driving and safe mobility, safe and secure wireless, telematics and multi-standard communication solutions, and the connected consumer, always with the guiding principle to identify and address highly valued use-cases and user-needs with differentiating technologies, tested and demonstrated in practical trial-environments. As such he has also participated in numerous national and transnational cooperation projects.
14:30-14:50
Presentation: C03_4: Connectivity and Cognitive Communication for Future Shared Mobility.
Gerald Fritz-Mayer, TTTech Auto AG, Austria.
The talk presents the current developments w.r.t. cognitive communication which enables AI-powered methods to access data from both the edge and the cloud, providing fast, reliable, and low-latency data connections. The presentation will showcase the respective demonstrators of (a) proof of concept communication platform and (b) novel technologies to prove wireless data transmission on 28GHz mmW frequency.
Speaker: Gerald Fritz-Mayer, TTTech Auto AG, Austria.
Gerald Fritz-Mayer (male) graduated in computer science at the Technical University of Graz and has an MBA degree on Entrepreneurship and Innovation from the Vienna University of Economics and Business. He joined TTTech as senior coordinator R&D project 2017 and is concerned with autonomous system in the field of manufacturing and autonomous driving. Beside that he participated in several national and European research projects, he was coordinating the innovation in various national research projects in the field of assistive system for future manufacturing systems. Gerald Fritz-Mayer is an IEEE member and serves in different officer roles in the robotic chapter Austria (RAS) and the Technology and Engineering Management Society (TEMS) Austria.
Presentation: D03_4: Accelerating Magnetic Resonance Imaging on Heterogeneous Edge Compute Platforms.
Jan Langer, Fraunhofer ENAS, Germany.
Simulating magnetic resonance imaging (MRI) is a computationally expensive task, that is based on solving the Bloch equations for every voxel and time step. Ideally, the simulation runs in real time on premise with confidential customer data, hence requiring a powerful edge acceleration platform. We developed a fully software-programmable and scalable solution for FPGA platforms achieving multiple magnitudes of performance increase. In this talk we will illustrate the opportunities and challenges we faced during this process.
Speaker: Jan Langer, Fraunhofer ENAS, Germany.
Jan Langer (MS’05 and PhD’11 in Electrical Engineering) leads the artificial intelligence team at Fraunhofer ENAS. Before that, he worked as a senior researcher and processor architect at Xilinx in the heterogeneous multi-core architectures team which developed the AI Engine. His current research at ENAS includes creating hybrid digital twins for semiconductor processes, as well as accelerating algorithms on edge AI platforms.
15:00-15:30
Afetrnoon Coffee/Tea Break
15:30-17:00
Edge AI Cluster Internal Meeting
17:00
End
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-17:00
Poster Session: 17-19 October, 2023
Poster: P01: Designing Trustworthy and Useful AI Applications: The Role of Human Oversight.
Nikolai Ebinger, Peter Moertl, Michael Karner, Virtual Vehicle Research GmbH, Austria.
AI is revolutionizing entire industries, but it cannot completely replace humans in many application areas. However, the human role is often not adequately considered in the design process, resulting in AI applications that challenge human operators to effectively oversee them – leading to distrust, undesired outcomes, and abandonment of the AI applications. To address this chain of distrust, we will present a model for human oversight of AI that builds on the experiences of the InSecTT project and sheds more light on the process behind and requirements for effective human-AI collaboration.
Poster: P02: Human acceptance and AI based Digital Workspaces.
Markus Tauber, Research Studios Austria, Austria.
Acceptance of AI in industry and e.g., technology uptake depends on ethics, legal norms, and trust. To support this acceptance, we are working on the production of guidelines and self-assessment approaches to support the evaluation of compliance with given standards and norms of AI based industrial solutions. A first initial step is the documentation of the existing landscape and stakeholder environment and requirements under consideration of the increasing popularity of technologies such as Large Language Models.
Poster: P03: Digital Industry.
Silke Braun, Josef-Anton Moser, Infineon, Austria.
Digitised control systems with advanced sensing, automated defect classification, virtual metrology and the implementation of a digital tacit knowledge platform are the clear vision for semi- autonomous AI supported production environment for a sustainable and resilient European Digital Industry. The value chain integrates the edge AI technological developments into six demonstrators.
Poster: P04: Energy Industry.
Ovidiu Vermesan, SINTEF AS, Norway.
Energy value chain develops advanced distributed edge AI technologies for applications and services at the intersection of mobility, energy and building industrial sectors to support the European stakeholders driving the engine for innovation and entrepreneurship in sustainable energy. The value chain integrates the edge AI technological developments into four demonstrators.
Poster: P05: Agri-food and Beverage.
Marcello Coppola, STMicroelectronics, France.
Agri-food and Beverage value chain develops technologies, procedures, and AI algorithms to make European agri-food and beverage industry globally renowned for its leadership, resilience, and contribution to climate neutrality. The value chain integrates the edge AI technological developments into four demonstrators.
Poster: P06: Mobility Industry.
Janis Arents, EDI, Latvia.
Mobility value chain develops edge AI-based perception, cognition, and monitoring technologies to advance the mobility sector in three different aspects: the mobile agent (e.g., drones, UAVs, UGVs and other vehicles), stationary and mobile multi-agent collaboration (distributed edge intelligence), and infrastructure. These technologies are key to enable connected and automated mobility with increased energy-efficiency, reliability, privacy, and reusability for both indoor and outdoor applications. The value chain integrates the edge AI technological developments into three demonstrators.
Poster: P07: Digital Society.
Gerardo Daalderop, NXP Semiconductor, the Netherlands.
Digital society requires the closer integration of technological systems and human users. This requires those systems to accurately interpret real world stimuli, inferring activity and intention. Such stimuli are poorly differentiated, indistinct, and contradictory. AI’s promise is to provide interpretation of stimuli in these situations that cannot be adequately modelled in the traditional way, thus allowing reactive processes to be better informed and enabling a closer integration of humans and digital infrastructure. The value chain integrates the edge AI technological developments into three demonstrators.
Poster: P08: Non-Volatile Memory (NVM) based In-Memory Computing (IMC) solutions of ANDANTE project.
Alexandre Valentian, CEA, France, Mario Diaz Nava, STMicroelectronics, France.
In data-centric applications, such as AI, memory accesses consume a large part of the total energy dissipation. This is why doing computations in-memory is one of the key solutions for reducing power at the edge. Furthermore, as the number of neural network parameters increases over time, relying on dense non-volatile embedded memories avoids duplicating memories and loading a considerable number of weights at each start up. Results shown here illustrate analog In-Memory Computing solutions, leveraging respectively MRAM, OxRAM and FeFET technologies.
Poster: P09: Edge vs. Cloud Implementation of Object Recognition in the Fridge with Camera Use Case.
Ebrar Ömer, Arcelik, Turkey.
In the Arçelik CamFridge smart refrigerator project, two cameras are strategically positioned one on the door and the other on the body This setup enables remote access for users to view refrigerator contents from their phones and explore eco friendly purchasing alternatives.
Poster: P10: EcoMobility: For a Sustainable Society with Intelligent Mobility Solutions.
Mohammed Abuteir, TTTechAuto, Austria Vicky Chatzidogiannaki, Innovation Dis.Co, Greece.
EcoMobility will develop and demonstrate a data driven framework for connected/ electric vehicles to enable a digital mobility ecosystem. The vision of the project is a sustainable door to door mobility based on autonomous vehicles with agile lifecycle mangement for improved safety, efficiency and ecology.
Poster: P11: Preparing corporates for the challenges of edge computing.
Jose Enrique Alvarez, BluSpecs, Spain.
With increasing computing and processing needs, edge computing will join cloud computing as the next important part of digital infrastructures. The convergence of Cloud and Edge will serve as an enabler of future developments. The growing need for edge computing solutions and building CEI networks is reflected in the analysis of expected spending on enterprise edge, estimated to double in the upcoming year. However, while industries and companies see CEI technologies’ potential for their business, not all of them are equally prepared or have the resources or skills to implement them successfully. Knowing how prepared they are to do so is critical to taking the proper steps to migrate to Edge Computing.
Poster: P12: Edge AI project.
Athanasia Thraskia, ITML, Greece, Teresa Koniari, ITML, Greece.
The poster gives an overview of the EdgeAI – Edge AI Technologies for Optimised Performance Embedded Processing, Key Digital Technologies (KDT) Joint Undertaking (JU) project, which is a key initiative for the European digital transition towards intelligent processing solutions at the edge. EdgeAI develops new electronic components and systems, processing architectures, connectivity, software, algorithms, and middleware by combining microelectronics, AI, embedded systems, and edge computing.
Poster: P13: Intelligent Motion Control under Industry4.E.
Vina Rompoti, GNT, Greece, Ilektra Simitsi, ITML, Greece, Athanasia Thraskia, ITML, Greece.
IMOCO4.E focuses on delivering an edge-to-cloud intelligence platform for cyber-physical systems, emphasizing digital twins and AI principles. This platform aims to enhance energy efficiency, reconfigurability, traceability, and cybersecurity in manufacturing applications, particularly in semiconductors, packaging, robotics, and healthcare, with potential applications in other motion control domains. IMOCO4.E’s impact extends across the production automation value chain.
Poster: P14: HW Design of Edge AI processors implementing Spiking and Artificial Neural Networks using different embedded Non-Volatile Memory technologies. Mario Diaz Nava, STMicroelectronics, France, Loreto Mateu Saez, Fraunhofer/IIS, Germany, Ivan Miro-Panades, CEA, France.
Optimal Edge AI solutions require high energy efficiency, low latency, and high integration. In the ANDANTE project, 14 circuits have been implemented in different novel eNVM (RRAM/OXRAM, FeFET and PCM) and FPGAs technologies. These designs cover different Neural Networks architectures (spiking, artificial and analogue). These circuits are being validated in several application domains: Digital Industry, Digital Farming, Transport and Smart Mobility, Healthcare and Digital life. This poster presents the circuits’ architecture and the main characteristics of each of them.
17:00
End
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session II: European Edge AI Ecosystem
9:00-9:30
Keynote presentation II_1
9:30-10:00
Keynote presentation II_2
9:00-09:30
Chips for Europe Initiative. Marco Ceccarelli, European Commission, Belgium.
The Chips Act allocates over 11 billion EUR of public support for the “Chips for Europe” Initiative. This initiative is structured around five primary lines of action: 1) a cloud-based virtual design platform, 2) pilot lines dedicated to prototyping and validation, 3) tools and infrastructure for quantum chips, 4) skills development and competence centres, 5) the Chips Fund, providing loans and equity investments. The envisaged virtual design platform will offer streamlined cloud access to tools, libraries, and services, targeting faster development and shorter time-to-market. Drawing from EUROPRACTICE’s success, the platform will provide EDA & IP, training, IC services, and prototyping for European enterprises. It aims to reduce barriers for SMEs, stimulate IP generation and sharing, and boost innovation. Users will get access to novel pilot lines, complete with PDKs, allowing them to develop prototypes using cutting-edge technologies and advanced packaging. The pilot lines are rooted in the testing and experimentation platform for Edge AI. This multi-hub project commenced at the end of 2022 and will offer prototype chip fabrication capability in advanced technologies for edge AI. The design platform will feature collaborative spaces to foster enhanced stakeholder cooperation in developing European technology, IP, and tools. Emphasis will be placed on the development of open-source hardware and tools. Edge AI is anticipated as a central theme for projects supported by the platform. The aim is to harness EU synergies in this arena, as edge AI can enable advancements in semiconductor technology nodes, design approaches, and architectural innovations.
Speaker: Marco Ceccarelli, European Commission, Belgium.
Marco Ceccarelli is Programme Officer at the European Commission in Brussels (B) where he contributes to the definition of EU policies for the electronics industry. He is one of the main authors of the European “Chips Act” package and plays a key role in its implementation. Before joining the Commission in 2017, he held several managerial roles at Philips (NL), where he focused on innovation for electronic systems, as well as several executive positions in high-tech SMEs. He began his career as research scientist in Italy and the Netherlands. Marco holds a Laurea degree in Electronic Engineering from the University of Florence (I) and an Executive MBA from the ESSEC business school in Paris (F).
9:30-10:00
The Landscape of Edge AI technology developments in Europe and the synergies with the EU Chips Initiative. Huascar Espinoza, Chips JU, Belgium
The long-term strategy of Chips JU is to tackle edge AI research and innovation by a series of open calls. While many ECSEL projects have been already contributing to the areas of edge computing, AI, IoT and connectivity, Chips JU has recently open edge AI-related Focus Topic calls oriented to push the technological boundaries in terms of processing solutions for AI at the edge addressing the design stack and middleware (Call 2021) and integration of trustworthy edge AI technologies in complex heterogeneous components and systems (Call 2023). Most recently a focus topic on AI-base methods and tools for engineering efficiency has started to be prepared with a primary scope on Edge AI. This talk provides a landscape of the Edge AI technology developments, the expected contributions to co-create a strong European edge AI ecosystem, and the synergies with the upcoming EU Chips Initiative particularly in liaison with the Chips Design Platform.
Speaker: Huascar Espinoza, Chips JU, Belgium.
Huascar Espinoza is a Programme Officer at Chips JU. He has a PhD in Computer Science, and his main background is on design of AI-based safety-critical systems. Prior to Chips JU, he acted as coordinator of various EU projects in the areas of safety assurance, certification, and robotics technologies.
10:00-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A02: Strategy for the Development of Edge AI Ecosystem.
Session Co-chairs:
Inessa Seifert, EPoSS, Germany, Paolo Azzoni, Inside Industry Association, the Netherlands
The goal of the session is to identify a potential roadmap for the European research and innovation in the domain of edge AI. The roadmap will coin future focus topics of the Chips Joint Undertaking and inspire other European programmes. We strive for an open ecosystem that fosters a fast adoption of research and innovation results in the European and global markets. We aim at strengthening the role of European chip vendors and system integrators in the emerging Digital Continuum and ensure high value proposition of key applications based on edge AI.
Keywords: edge AI ecosystem, European Chips Act, digital continuum, research and innovation.
Session B02: Smart Connectivity at the Edge.
Session Co-chairs:
Luca, Scuola Superiore Sant’Anna, Italy, Marcello Coppola, STMicroelectronics, France.
Edge systems will increasingly leverage collaboration to reach computational effectiveness while reducing hardware computational requirements (e.g., swarm learning). In addition, in network computation is emerging to reduce capacity requirements and closed-loop control execution time. This scenario demands for smart connectivity, both wired and wireless, at the edge where resources are allocated for data and control communications preserving energy, capacity, security and confidentiality. This session will focus on tools and techniques for providing the requested connectivity to the edge devices while guaranteeing the KPIs.
Keywords: wireless, wired, connectivity, collaborative computing, energy, capacity, security, confidentiality.
11:00-11:20
A02_1: Edge AI Stakeholders Covering the Value Chain
Presentation and discussion on the edge AI value chain that encompasses various stakeholders who contribute to the development, deployment, and utilization of edge AI technologies. Alignment among these stakeholders is important to realizing the full potential of edge AI while addressing challenges related to security, privacy, interoperability, and ethical considerations.
Presentation: B02_1: Edge AI LoRa Mesh Technologies.
Janis Judvaitis, Institute of Electronics and Computer Science, Latvia.
Intelligent connectivity at the edge combines wireless communication, edge artificial intelligence (AI), edge computing and Internet of Things (IoT) technologies to perform machine learning (ML) and deep learning (DL) tasks on connected edge devices. Low latency, ultra-low energy intelligent IoT devices with on-board computing, distributed architecture and analytics are essential as intelligent connectivity advances.
Speaker: Janis Judvaitis, Institute of Electronics and Computer Science, Latvia.
Janis Judvaitis is a researcher in Institute of Electronics and Computer Science (EDI) as well as a Ph.D student in Computer Science at University of Latvia. Works in EDI from 2013. He is involved in various international projects related to Wireless Sensor Networks, embedded systems, and Internet of Things. Expertise in ICT system design, development, implementation, and testing.
11:20-11:40
A02_2: European Ecosystem and Synergies for Fast Adoption of Edge AI Solutions.
Presentation and discussion on the fast adoption of edge AI solutions, the development of a robust ecosystem and the creation of synergies among various stakeholders. Focus on the key elements and strategies for fostering this ecosystem and accelerating the adoption of edge AI.
Presentation: B02_2: Cascaded Look Up Table Distillation of P4 Deep Neural Network Switches.
Luca Valcarenghi, Scuola Superiore Sant’Anna, Pisa, Italy.
In-network function offloading represents a key enabler of the SDN-based data plane programmability to enhance network operation and awareness while speeding up applications and reducing the energy footprint. This paper proposes an innovative knowledge distillation technique that maps a DNN into a cascade of lookup tables (i.e., flow tables) with limited entry size. The proposed mapping avoids stateful elements and maths operators, whose requirement prevented the deployment of DNNs within hardware switches up to now. The evaluation considers a cyber security use case targeting a DDoS mitigator network function, showing negligible impact due to the lossless mapping reduction and feature quantization.
Speaker: Luca Valcarenghi, Scuola Superiore Sant’Anna, Pisa, Italy.
Luca Valcarenghi is an Associate Professor at the Scuola Superiore Sant’Anna of Pisa, Italy, since 2014. He received the Laurea in Electrical Engineering in 1997 from Politecnico di Torino and the M.S.E.E. and Ph.D. in Electrical Engineering Major Telecommunications from UTD in 1999 and 2001, respectively. He published more than three hundred papers in International Journals and Conference Proceedings. Dr. Valcarenghi received a Fulbright Research Scholar Fellowship in 2009 and a JSPS “Invitation Fellowship Program for Research in Japan (Long Term)” in 2013. His main research interests are optical networks design, analysis, and optimization; communication networks reliability; energy efficiency in communications networks; optical access networks; zero touch network and service management; 5G technologies and beyond.
11:40-12:00
A02_3: Collaboration Along the AI Value Chain
Collaboration along the AI value chain is crucial for the development, deployment, and success of AI solutions. The discussion will focus on which are the AI value chain various stages, from data collection and preprocessing to model development, deployment, ongoing optimization and what are the catalysts for increasing the collaboration.
Presentation: B02_3: Energy Optimisation for Edge Devices.
Javad Chamanara, University of Hannover, Germany.
The presentation gives an overview of RLKube, an RL-based custom Kubernetes scheduler plugin for optimized task scheduling. RLKube’s objective is to improving energy efficiency in a Kubernetes cluster. RLKube considers multiple optimization objectives, including energy optimization and Pod throughput. It utilizes DDQN with PER with different reward functions to train the RL agent. Our evaluations indicate the effectiveness of RLKube in outperforming the default Kubernetes scheduling policies in terms of throughput and energy usage.
Speaker: Javad Chamanara, University of Hannover, Germany.
Dr. Eng. Javad Chamanara has a PhD computer science. He is a research team lead at the University of Hannover Germany. His research projects are focused on federated learning, energy optimization, and privacy-aware computing. He is a participating to the CLEVER, BRAINE, and Glaciation projects to provide privacy-aware, energy optimization to federated data operations.
12:00-12:20
A02_4: Market Impact Along AI Value Chain for Co-creation of Edge Applications.
The discussions will focus on the co-creation of edge applications, where multiple stakeholders collaborate along the AI value chain to develop and deploy edge AI solutions, and how this co-creation have significant market impacts.
Presentation: B02_4: Edge-assisted Smart Agriculture for High Yield Eco-farms.
Alessandro Pacini, Scuola Superiore Sant’Anna, Italy.
This presentation describes the architecture and lists a preliminary set of requirements of an edge-assisted smart agriculture application for high yield eco-farms. In particular, requirements in terms of computational resources, energy, latency, transmission capacity are considered in the specific deployment of the CLEVER project.
Speaker: Alessandro Pacini, Scuola Superiore Sant’Anna, Italy.
Alessandro Pacini is a PhD student in Emerging Digital Technologies at the Scuola Superiore Sant’Anna. He received his bachelor’s degree in computer science from the University of Camerino in 2018. He then pursued his studies and obtained a joint master’s degree in computer science and Networking at the University of Pisa and Scuola Superiore Sant’Anna in 2021. During this time, he won a one-year research fellowship at SSSA focused on building a scalable and reliable monitoring architecture for optical networks. His main research interests are in the field of next-generation SD networks, with a particular focus on reusing existing network architectures to move towards a zero-touch paradigm.
12:30-13:30
Networking Lunch
13:30-15:30
Session C02: Edge AI Technology Development.
Session Co-chairs:
Reiner John, AVL LIST Gmbh, Austria.
The session will address the convergence of technologies and ecosystems in different industrial sectors and how using AI and standardisation combined with dynamic change reshapes industries, breaking down silos and fostering cross-sectoral collaboration.
Keywords: edge AI technologies, industrial sectors, technology-driven synergy, catalysts for innovation.
Session D02: Agri-Food and Beverage.
Session Co-chairs:
Markus Tauber, Research Studios Austria, Austria, Aymeric Deshayes, Laboratoire IMS, France
AI at the edge, on devices located closer to the data source, such as sensors and machinery within the agricultural and food processing sectors is advancing rapidly. It brings intelligence and automation to various processes, optimising efficiency, reducing waste, improving quality, and ensuring sustainability. The session covers several innovative edge AI applications in the domain with the goal of improving sustainability, reducing waste, enhancing quality, and ensuring agr-food and beverage safety throughout the entire supply chain.
Keywords: precision agriculture, energy efficiency, safety and traceability, resource optimisation, real-time decision support, yield prediction, disease identification, crop production optimisation.
13:30-13:50
Presentation: C02_1 – Cutting-edge Technology Development, and Ecosystems Convergence.
Reiner John, AVL LIST Gmbh, Austria.
In the realm of cutting-edge technology, the concept of ecosystem convergence takes center stage. This speech dives into the intricate interplay of industries and technologies, driving collaboration and innovation in the quest for transformative solutions, even if specific initiatives remain undisclosed. Ecosystem convergence signifies the merging of traditionally separate sectors, ushering in new synergies and unprecedented opportunities. We embark on an exploration of how this dynamic phenomenon reshapes industries, breaking down silos and fostering cross-sectoral collaboration. The evolving ecosystem is marked by its fluidity, where industries seamlessly collaborate across boundaries. We’ll delve into real-world examples of how convergence serves as a catalyst for innovation, revolutionizing fields such as smart manufacturing, sustainability practices, and beyond. At the heart of this transformation lies Artificial Intelligence, an omnipresent force driving synergy within the ecosystem. As AI permeates diverse sectors, its transformative influence aligns seamlessly with the pursuit of efficiency, digitalization, and software-defined transformations. Join us as we navigate the uncharted terrain of ecosystem convergence, propelled by a commitment to collaboration and innovation. Together, we will unlock the full potential of technology-driven synergy, reshaping the landscape of industries and technologies.
Speaker: Reiner John, AVL LIST Gmbh, Austria.
Reiner John, an accomplished electrical engineer, began his career at Siemens Semiconductor Group before transitioning to Siemens Automation Group, where he specialized in embedded control development tools. His expertise expanded into knowledge-based embedded systems and automotive systems research at Siemens’ Central Research and Development. He later held management positions at Infineon Technologies and established their Silicon Foundry Taiwan Office. Since 2007, Reiner has been instrumental in initiating and coordinating publicly funded projects. Currently, as the Coordinator of Research Funding in Corporate Strategy at AVL Graz, he focuses on advancing electromobility and driving innovation in cutting-edge technologies and AI-driven digitalization.
Presentation: D02_1: Using Edge AI in IoT Devices for Smart Agriculture Applications: Autonomous Weeding of Vegetables and Vine Health Monitoring.
Aymeric Deshayes, Laboratoire IMS, France.
Crop production uses increasingly fixed sensors in the plots, on-board sensors on tractors and robotic machines. These devices frequently have limited access to networks which make EDGE architectures particularly suitable. Through two real life use cases, this article shows how applications based on artificial vision can benefit from electronic circuits, making it possible to use artificial intelligence in low-cost EDGE architectures.
Speaker: Aymeric Deshayes, Laboratoire IMS, France.
Aymeric Deshaye is a research engineer who graduated in 2017 from the agronomic school of Bordeaux Sciences Agro, with a specialization in digital for agriculture. He is currently working at the IMS research Lab. (Bordeaux – France) in a group working more specifically on image processing applied to agriculture. Aymeric Deshayes is involved in various research projects including vine disease detection, robotic weeding and Edge Computing for Crop Monitoring. He is currently project manager for IMS Lab within the ANDANTE Project.
13:50-14:10
Presentation: C02_2: Intelligent, Safe, and Secure Connected Electrical Mobility Solutions.
Mohammed Abuteir, TTTech, Austria.
The mission of EcoMobility is to support European industry, transportation infrastructures and cities in transitioning from isolated and static transportation means towards a service-centric, connected mobility ecosystem by sharing data across involved stakeholders. The presentation will focus on the AI for edge/cloud-based service deployment, monitoring, analysis and coordination of vehicles, transportation infrastructures, people and goods.
Speaker: Mohammed Abuteir, TTTech, Austria.
Dr Mohammed Abuteir holds a PhD in Embedded Systems from the University of Siegen, Germany. His research focused on “Architecture Design for Distributed Mixed Criticality Systems based on Multi-Core Network on a Chip” with an emphasis on system architectures, reliability and adaptivity for distributed embedded real-time systems. He has also researched simulation environments for time-triggered communication. In 2017, Dr Abuteir joined TTTech, where he is involved in the management of large R&D projects in the field of safety-critical and autonomous systems, particularly in the field of autonomous driving. In his role, he coordinates and manages several projects such as the EcoMobility project and the AutoDevSafeOps project. He has also been project manager for several EC projects such as SYS2WHEEL, SAFE4RAIL-3, Up2Date, productive 4.0 and AutoDrive.
Presentation: D02_2: Vineyard Yield Prediction.
Lilian Hollard, Université de Reims Champagne-Ardenne, France.
The Value Chain Agri-food and Beverage initiative strives to equip the European Agri-food and Beverage industry with pioneering technologies, protocols, and AI algorithms. This endeavour seeks to establish the industry’s global eminence through its leadership, resilience, and contributions to achieving climate neutrality. Within the framework of this project, we showcase the latest advancements in technology and applications, with a particular emphasis on AI and edge AI methodologies. These methodologies and techniques, used predominantly in the context of vineyard yield forecasting, stand out in the yield forecasting demonstrator, which is dedicated to crafting a dependable edge AI framework for precise yield forecasting, striving for an error margin of less than 15%.
Speaker: Lilian Hollard, Université de Reims Champagne-Ardenne, France.
First year Ph.D. student at LICIIS from the University of Reims Champagne Ardenne, development of Artificial Intelligence models for viticulture: Disease identification and yield estimation. Adaptation and optimization of Deep Learning models for Edge and mobile devices.
14:10-14:30
Presentation: C02_3: PREVAIL project: Test and Experimentation Facility for the Realization and Validation of Prototype Chips Dedicated to Edge AI.
Sergio Nicoletti, CEA-Leti, France.
The PREVAIL project will establish and start operating the core of a networked, multi-hub platform providing prototype chip fabrication capability, in advanced to EU stakeholders for Artificial Intelligence (AI) applications. In this project, four major European RTOs jointly work to create a multihub Test and Experimentation Facility for edge AI Hardware (TEF HW AI) build on their advanced 300 mm fabrication, design, and test facilities. Within PREVAIL, companies and research institutions will be able to develop, test and experiment with AI product prototypes based on innovative technologies developed by the four RTOs. The technologies made available are the most promising to address the specifications (power consumption, computing performance, data handling and security) required by AI chips in general and edge AI in particular.
Speaker: Sergio Nicoletti, CEA-Leti, France.
Sergio Nicoletti Ph.D joined the CEA in 2006 to develop optical gas sensors. Nicoletti has since served as chief researcher and project leader in several industrial and collaborative projects. Currently, he coordinates the TEF-Prevail Project. He earned a doctorate in physics from Université Joseph Fourier in Grenoble, France.
Presentation: D02_3: Disease Identification for Vineyards Fields.
Lucas Mohimont, Université de Reims Champagne-Ardenne, France.
Disease Identification aims to progress with an advanced edge AI framework tailored for real-time disease detection in the field. The latest techniques are implemented within an object detection architecture, effectively enhancing the deep learning model for tomato disease prediction. This enhancement enables real-time, on-site predictions, facilitating efficient large-scale processing. In the initial stages, meta-edge analysis could contribute to generating a map of affected zones, offering champagne manufacturers and winegrowers the means to optimize the application of phytosanitary products. The presentation focuses on state-of-the-art technological and application developments with a spotlight on AI and edge AI methods and techniques used in the project.
Speaker: Lucas Mohimont, Université de Reims Champagne-Ardenne, France.
Ph.D. in Artificial intelligence for industry. Application to viticulture and COVID-19 case modelling. Lucas Mohimont is working at the University of Reims Champagne Ardenne for the development of Artificial Intelligence models for viticulture: Disease identification and yield estimation.
14:30-14:50
Presentation: C02_4 – Edge AI Standardisation Activities.
Erwin Schoitsch, AIT Austrian Institute of Technology, Austria.
Presentation and discussions on standardisation activities, EU AI Act and the challenges related to verification, validation and testing of edge AI systems.
Standardization is key to a fair, competitive society and economy. This includes technical standards, specifications and reports developed by international, European and/or industrial alliances in generic and domain-specific areas of interest. Of public interest and for public acceptance of new technologies are issues concerning safety, dependability, cybersecurity and ethical concerns, particularly in context of decision making algorithms and Artificial Intelligence. On European level, the carbon-footprint and “Green Deal” contributions play an important role as well. The presentation will provide an overview on the current efforts in standardization and by publishing guidelines and recommendations, and how they are addressed in the project AI4CSM (Automotive Intelligence for Connected Shared Mobility).
Erwin Schoitsch works more than 50 years in different positions for AIT Austrian Institute of Technology. Since his retirement in 2010 he works as Free Lancer under various subcontracts in many European Research projects for AIT and other organizations. Focus of his engagement was always on trustworthiness of highly reliable, safety-critical systems (Dependability, Safety, Cybersecurity), in industrial projects as well as European Research. He is active in many standardization groups of ISO, IEC and ISO/IEC JTC1, and the Austrian National Mirror Committees, covering topics like functional safety, cybersecurity, automated driving, Artificial Intelligence and ethical concerns. He is in several European Research Projects leading the standardization tasks.
Presentation: D02_4: CamFridge: Edge vs. Cloud Implementation of Object Recognition in the Fridge with Camera Use Case.
Ebrar Ömer, Arcelik, Turkey.
This study addresses a pivotal choice: adopting edge (local devices) or cloud computing for our smart camera-equipped fridge. Cloud systems utilize remote servers, while edge computing processes data instantly on devices. We aim to assess the YOLOv7 model’s efficacy in both paradigms for our smart fridge, given their distinct merits and challenges.
Speaker: Ebrar Ömer, Arcelik, Turkey.
Ebrar Ömer, is a 23-year-old computer engineering graduate from Istanbul Technical University (İTÜ). Currently, she works at Arçelik and has a deep passion for artificial intelligence and machine learning. Fluent in both Turkish and English, she aspires to build her career in these dynamic fields, contributing to the ongoing transformation of the technology landscape.
15:00-15:30
Afternoon Coffee/Tea Break
15:30-17:00
Session E02: Energy Optimisation for Edge Devices.
Session Co-chairs:
Pavel Smrz, COGNITECHNA s.r.o., Czech Republic, Paolo Meloni, Università degli studi di Cagliari, Italy
This session is dedicated to the research on power-aware edge computing, which aims to optimize edge devices’ energy efficiency and performance by dynamically adapting to the workload, communication needs, and environmental conditions. Presented papers will discuss the benefits of running optimised AI-based solutions regarding decreased power requirements and long-life operation for near real-time data processing.
Keywords: power awareness, energy efficiency, micro-power management, long battery life, dynamic adaptation to workload, energy-optimised AI, machine learning for energy optimisation.
Session F02: Edge AI Hardware Platforms.
Session Co-chairs:
Björn Debaillie, imec, Belgium, Olaf Peters, Technolution, the Netherlands.
Edge AI hardware platforms are specialised computing systems designed to accelerate and optimise the execution of AI workloads on edge devices. Various platforms are used for running AI models efficiently at the edge, where resource constraints, power efficiency, and real-time processing requirements. The session gives an overview of different technological developments addressing several edge AI HW platforms.
Keywords: multi-application HW platforms, FPGAs, Soc, SoM, MCUs, heterogeneous integration, SNN processor platforms, automated optical inspection edge AI inference platforms, heterogeneous edge systems, platforms for federated learning.
15:30-15:50
Presentation: E02_1: Tiny Transformers for On-line Processing of Sensor Data Streams on Low-power MCUs.
Paola Busia, Università degli Studi di Cagliari, Italy.
The presentation discusses the design, training and deployment of lightweight energy efficient algorithms derived from vision transformers exploited for the real-time analysis of sensor data streams. The capabilities of the presented approach are assessed on decoding and classification tasks involving biosignals. The algorithms are executed on several parallel MCU for IoT-node development.
Speaker: Paola Busia, Università degli Studi di Cagliari, Italy.
Paola Busia is assistant professor at Università degli studi di Cagliari. Her main research interests involve power efficient processing architectures for embedded applications and for at-the-edge AI.
Presentation: F02_1: SNNs on FPGAs for On-line Processing of Sensor Data.
Paolo Meloni, Università degli Studi di Cagliari, Italy.
The presentation discusses the design, training and deployment of lightweight energy efficient spiking neural networks exploited for the real-time analysis of sensor data streams. The capabilities of the presented approach are assessed on decoding and classification tasks involving biosignals. The algorithms are executed on a custom SNN processor deployed on low/mid-end FPGAs.
Speaker: Paolo Meloni, Università degli Studi di Cagliari, Italy.
Paolo Meloni is associate professor at Università degli studi di Cagliari. His main research interests involve power efficient processing architectures for embedded applications and for at-the-edge AI.
15:50-16:10
Presentation: E02_2: CA-CFAR is Convolution: Fast Target Detection With Machine Learning Accelerator.
Chen Liu, Technische Universität Dresden, Germany.
In radar target detection, Constant False Alarm Rate (CFAR) is a commonly employed detector known for its simplicity and effectiveness. Its sliding-window detection mechanism possesses computational similarity to convolutional operations in machine learning. With the increasing emergence of AI-enhanced radar processing algorithms, systems at the edge are often equipped with ML accelerators to expedite matrix multiplications and convolutions. This paper introduces a heuristic algorithm that equivalently maps Cell-Averaging (CA) CFAR to a convolutional operation and leverages machine learning accelerators for acceleration. Comparative experiments against the conventional CA-CFAR algorithm executed on ARM cores demonstrate that our proposed method significantly reduces the processing latency by 38 times without incurring substantial memory overhead. This advancement brings substantial promise for potentiating real-time high-resolution radar target detection.
Speaker: Chen Liu, Technische Universität Dresden, Germany.
Chen Liu received the BSc degree and MSc degree in electrical engineering in 2012 and 2016, respectively. He is currently working toward the PhD degree in Technische Universität Dresden, Germany. His research interests include intelligent radar signal processing, efficient deep learning on the edge and neuromorphic computing in automotive.
Presentation: F02_2: ANDANTE Platform 4.1a : A Multi-Application Platform Supporting Several Uses Cases in the Domains Digital Farming and Transport and Smart Mobility.
Ivan Miro-Panades, CEA, France.
An IoT challenge is to provide devices supporting various applications to reduce cost and parts management. As part of ANDANTE, a multi-application platform has been developed to address 6 use cases in two different applications domains Digital Farming and Transport and Smart Mobility. Platform 4.1a is a combination of an AI ASIC accelerator, NeuroCorgi, and a commercial FPGA including an ARM core. The implementation of this platform required to make a compromise between the FPGA and the ASIC to share the ANN layers and to optimise them according to the needs of the application to guarantee high application performance (energy efficiency, latency, accuracy, and good level of integration).
Speaker: Ivan Miro-Panades, CEA, France.
Ivan MIRO-PANADES received the M.S. degree in telecommunication engineering from the Technical University of Catalonia (UPC), Barcelona, Spain, in 2002, and the M.S. and Ph.D. degrees in computer science from Pierre and Marie Curie University (UPMC), Paris, France, in 2004 and 2008, respectively. He worked at Philips Research, Sureness, France and STMicroelectronics, Crolles, France, before joining CEA, Grenoble, France, in 2008, where he is currently an Expert Research Engineer in digital integrated circuits. His main research interests are artificial intelligence, the Internet of Things, low-power architectures, energy-efficient systems, and Fmax/Vmin tracking methodologies.
16:10-16:30
Presentation: E02_3: Radar-based Reactive Planning for Drones.
Hamed Javadi, imec, Belgium.
While cameras and lidars are more common in reactive planning in drones, radars can potentially complement, or replace, them thanks to their robustness against harsh weather and extreme light conditions. To this end, appropriate signal processing algorithms are needed to tackle the poor angular resolution of radars. These algorithms, on the other hand, should fit the limitations at edge (drone). In this presentation, we present a computationally efficient synthetic aperture radar (SAR) pipeline as a solution to radar-based reactive planning in drones.
Speaker:Hamed Javadi, imec, Belgium.
S. Hamed Javadi is working as a researcher at imec, Belgium, in the fields of signal processing, sensor fusion, and AI. His current work focuses on the development of AI algorithms for radar signals which also requires strong radar signal processing as a preprocessing step. Before joining imec, S. Hamed Javadi was a postdoc at Ghent University, Ghent, Belgium, where he developed efficient AI-based algorithms for smart farming based on multiple sensors’ data. S. Hamed Javadi was also an assistant professor at the University of Bojnord, Iran, where his main focuses of interest were statistical signal processing, distributed algorithms, and machine learning.
Presentation: F02_3: Automated Optical Inspection Edge AI Inference System Solutions.
Claudio Cantone, High Technology Systems H.T.S. srl, Italy and Alberto Faro, Deepsensing, DEEPS, Italy.
The presentation gives an overview of the automated optical inspection edge AI inference system solutions in digital industry by discussing if and how they allow the manufacturers to achieve a satisfactory trade-off between customer requirements and production costs. The comparison of the available solutions using KPIs, FRs and NFRs points out that AOI consisting of testing boards at the edge based on a model pre-trained on the cloud server is a promising solution if it is provided with a suitable worker interface and planning tools for managers.
Speaker: Claudio Cantone, High Technology Systems H.T.S. srl, Italy and Alberto Faro, Deepsensing, DEEPS, Italy.
Claudio Cantone received the Electronic Engineering degree from University of Pisa and the Master in Business Administration from Bocconi University. Initially he worked for several years at Philips Spa, Cinisello Balsamo, Milan in the sector of Electromedical Technical Services. Then he founded HTS srl that is a company providing Electronic Manufacturing Services (EMS) to support the entire life cycle of products in digital industry. Currently he is the Head of the HTS R&D Department.
Alberto Faro received an engineering degree from Milan Polytechnic. He was one the scientists who designed and tested the first European Informatics Network (EIN), forerunner of the Internet. From 1980 to 2016 he was Professor of Artificial Intelligence at the University of Catania where he contributed to find the Electronic and Computer Engineering Degrees. In 2019 he founded Deepsensing srl that is an innovative Startup active in the Artificial Intelligence of Things (AIoT) field.
16:30-16:50
Presentation: E02_4: Inside the AI Accelerators: From High Performance to Energy Efficiency.
Ana Pinzari, Univ. Grenoble Alpes, France.
This presentation covers current technologies for high-performance, low-power neural networks. To cope with the high computational and storage resources, hardware optimisation techniques are proposed: DL compilers and frameworks, DL hardware and hardware-specific code generator (FPGA). More specifically, in this presentation we explore the quantization mechanism in deep learning, based on a deep-CNN classification model. We highlight the accuracy of quantized models and explore their efficiency on a variety of hardware platforms. Through experiments, we show the performance achieved using general-purpose hardware (CPU and GPU) and a custom ASIC (TPU), as well as the simulated performance for a reduced bit-width representation of 4 bits, 2 bits (ternary) up to 1-bit heterogeneous quantization (FPGA).
Speaker: Ana Pinzari, Univ. Grenoble Alpes, France.
Ana Pinzari received her Ph.D degree in computer science at Université de Technologie de Compiègne, Compiègne, France in 2012, in Real-Time Embedded Systems. Then, she worked in a scientific research organization ECSI (European Electronic Chips & Systems Initiative) focusing on new design methods, tools, and standards for design of complex electronic systems. In 2021, she joined Grenoble INP, TIMA, as a post-doctoral research engineer. Her current research focuses on exploring optimisation methods for the hardware implementation of neural networks.
Presentation: F02_4: Federated Learning in Cybersecurity.
Dimitrios Serpanos, CTI and University of Patras, Greece.
Federated learning constitutes a promising solution for ML/AI-enabled cybersecurity solutions in distributed environments with Edge systems. We present an efficient malware detection solution employing federated learning for environments that combine powerful as well as limited resource edge systems. We demonstrate that federated learning enables effective cybersecurity solutions for distributed environments that include heterogeneous edge systems and have constraints and limitations in data exchange, ranging from performance to IP.
Speaker: Dimitrios Serpanos, CTI and University of Patras, Greece.
Dimitrios Serpanos is president of the Computer Technology Institute & Press DIOPHANTUS and Professor of ECE at the University of Patras, Greece. He is Chair of the Scientific Council of INSIDE Industrial Association.
17:00
End
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session I: Edge AI Technology Advancements
9:00-9:30
Keynote presentation I_1
9:30-10:00
Making Edge AI a reality. Pierrick Autret, STMicroelectronics, France.
The emergence of Edge AI has opened new opportunities for smart devices and applications. However, the implementation of Edge AI requires low-power hardware and software platforms that can handle the processing demands of AI algorithms. In this talk STMicroelectronics will present how to address these challenges and provides solutions to make Edge AI a reality. The presentation will cover topics such as the latest developments in low-power hardware and software platforms, and how to optimize Edge AI performance while minimizing power consumption. Finally, the presentation will showcase some of the latest ST products and technologies that enable the development of Edge AI applications.
Presenter: Pierrick Autret, STMicroelectronics, France.
Pierrick Autret has been Artificial Intelligence Solutions Manager in STMicroelectronics’ microcontroller division since 2019. With a master’s degree in engineering specializing in Embedded Systems from the Ecole Polytechnique and a master’s degree in Marketing, Economics and Management from the IAE d’Aix-en-Provence, he began his career as Hardware Product Owner on the STM32 series before developing ST’s software offering.
10:00-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A01: Edge AI Tools and Methods.
Session Co-chairs:
Fabian Chersi, CEA-List, France, Frédéric Pétrot, Univ. Grenoble Alpes, France
This session is dedicated to the presentation of tools and methods aimed at simplifying and accelerating the design flow, from application development to deployment, from neural network exploration, learning, quantizing, testing, to the generation of optimized hardware-aware code and deployment on edge devices constrained by power consumption or dissipation, latency, form factor and/or cost criteria.
Keywords: edge devices, quantization, pruning, NN optimization, NN deployment, tools for edge devices.
Session B01: Edge AI Machine Vision.
Session Co-chairs:
Alain Pagani, DFKI, Germany, Sadique Sheik, SynSense AI, Switzerland
The session covers presentation addressing edge AI machine vision technology and applications developments based on different algorithms and techniques. The presenters focus on the optimisation techniques and AI accelerators capable of processing inference at minimal power consumption. The implementations include ML algorithms and neural networks used in machine vision tasks optimises for edge devices with constrained resources.
Keywords: edge AI, machine vision, machine learning, deep learning, image signal processors, convolutional neural networks.
11:00-11:20
Presentation: A01_1: Optimization and Compression of Computer Vision and Generative AI Models for Neuromorphic Accelerators.
Tim Llewellynn, Bonseyes Community Association, Switzerland.
In many AI applications, models typically demand extensive training datasets and benchmark data for robust performance. However, gathering such vast and diverse datasets can be especially challenging in edge computing environments. This presentation addresses this critical challenge by showcasing how Generative AI and Computer Vision foundational models can help mitigate the burden by generating synthetic data and benchmarks that closely resemble real-world scenarios. Applicability to Neuromorphic model development will be discussed and AI development in edge computing, where data scarcity is a common constraint.
Speaker: Tim Llewellynn, Bonseyes Community Association, Switzerland.
Tim Llewellynn is President of the Bonseyes Community Association and has 20 years in high-tech startups and corporate R&D in data-driven software development for cloud, mobile and edge computing applications. At Bonseyes Community Association (BCA), is a non-profit AI-focused association, he drives the democratization of artificial intelligence for the benefit of all society, ensuring that AI, from narrow to general, creates highly autonomous systems that benefit everyone, not just a select few. Tim received his B.E. and M.E. (Hons) degrees in Electrical Engineering from the University of Canterbury, New Zealand in 1999, and 2001.
Presentation: B01_1: Image Signal Processor (ISP)
Tunning Machine Learning (ML) methods.
Sepehr Bijani, NXP Semiconductors, Munich, Germany
Image Signal Processor (ISP) is responsible for improving camera signal quality and producing a high-quality image. ISP has a vast number of parameters which should be tunned based on both camera sensor and the environment which camera is operating. Sensor related ISP parameters should be tunned offline for each camera sensor. Parameters related to environment such as white balancing gains should be fine-tuned in the field during the runtime. The offline phase is cumbersome and costly. At the same time, the runtime for fine-tuning should be fast and accurate. Therefore, the Tuner needs to achieve two goals: a) tune the sensor related parameters automatically in lab and b) create a data set for different scenario in lab and train a Machine Learning (ML) runtime for fine-tunning ISP in the field. We first create a solver for finding static camera related parameters. Then, we create the dataset of camera scenes and optimal ISP parameters found by the same solver for each scene for training runtime ML model. The trained ML model will inference optimal ISP parameters for the scene in the field.
Speaker: Sepehr Bijani, NXP Semiconductors, Munich, Germany.
Sepehr is SW engineer working for NXP Germany. His main research topic is AI/ML in image processing applications. His professional carrier focuses on embedded SW engineering.
11:20-11:40
Presentation: A01_2: Shrinking Embedded Machine Learning Applications: FMCW Radar Gesture Sensing on Cortex-M0+ under 10KB and 10mW.
Jonas Daugalas, Infineon Technologies AG, Germany
AI-enabled millimeter-wave radar sensors are becoming increasingly popular solutions for human-machine interaction tasks. However, due to compute and memory requirements, signal processing and machine learning approaches pose challenges for deploying such solutions to low-cost and low-power edge devices. In this work we demonstrate signal processing optimization techniques and neural network inference engine trade-offs that allow us to deploy a 60GHz FMCW radar-based gesture recognition solution on an Infineon PSoC™ 4 microcontroller equipped with an Arm® Cortex-M0+ processor. Our proposed system robustly recognizes five distinct hand gestures, requires less than 10 milliwatts of power, 9.7KB of RAM, 107KB of EPROM, 73% of processor time, and with microTVM, among other three evaluated inference engines, achieves the best trade-off between manual effort and in-field performance.
Speaker: Jonas Daugalas, Infineon Technologies AG, Germany.
Jonas Daugalas has a background in Computer Science, bachelor’s degree in CS from Vilnius University and master’s degree in CS from Technical University of Munich. Software engineering experience in the fields of particle physics, computer-aided design, physical simulation, machine learning and digital signal processing.
Jianyu Zhao has background in Electrical and Information Engineering: received a B. Eng. from Tianjin University in 2015 and a M. Sc. from Technical University of Munich in 2018. Work experience on algorithm design and deployment for smart sensing technologies at Infineon Technologies AG since 2018. Current research interest: neural network embedding on edge devices.
Presentation: B01_2: Ultra-low Power Car Plate Recognition at the Edge.
Dávid Isztl, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
So far, only a few edge devices include AI accelerators that are capable of processing inference at minimal power consumption. As an example, the MAX78002 provides a 2Mbytes CNN engine and an Arm Cortex-M4 with 384KB of SRAM. On one hand, the device can work battery-powered for up to 3 years. On the other hand, severe compromises and complex procedures are necessary to adapt AI networks for specific applications. Using license plate recognition as an example, the talk will show how the shortcomings of the small memory footprint and low image resolution offered by the AI accelerator can be overcome.
Speaker: Dávid Isztl, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
Dávid Isztl is a research assistant at the Institute of Embedded Systems at ZHAW. Currently, his research is focused on high-performance embedded platforms, including AI accelerators and Nvidia NX/AGX platforms. Currently, he is focusing on low-power AI processing on the edge.
11:40-12:00
Presentation: A01_3: Aidge: A New Framework for Deep Neural Network development, Training, and Deployment on the Edge.
Fabian Chersi. CEA-List, FRANCE
Aidge is a comprehensive framework for fast and accurate Deep Neural Network (DNN) building, training, testing and deployment. The platform integrates tools for database construction, data pre-processing, network building, benchmarking, and hardware export to various targets. It is particularly useful for DNN design and exploration, allowing simple and fast prototyping of DNN with different topologies. Export hardware targets include CPU, DSP, and GPU as well as custom hardware code generation with High-Level Synthesis for FPGA and dedicated configurable DNN accelerator IP.
Speaker:Fabian Chersi. CEA-List, FRANCE.
Dr. Fabian Chersi has a BSC and a Master in Electronic Physics, and a PhD in Computational Neuroscience and Robotics. After his studies, he initially did a postdoc at the Institute for Neuroinformatics in Zurich and at UCL in London, then he moved to Paris where he worked in private companies on the development of neuromorphic algorithms and ASICs for deep neural networks. He currently works at the Alternative Energies Commission (CEA) in Paris as a senior AI architect on low power AI chips for the edge.
Presentation: B01_3: Designing Lightweight CNN for Images: Architectural Components and Techniques.
Lilian Hollard, Université de Reims Champagne-Ardenne, France
While neural networks have brought about impressive advancements in computer vision tasks, these achievements heavily depend on computationally demanding resources, restricting their deployment. The presentation unveils the landscape of deep learning architecture optimization tailored for mobile and Edge devices.
Speaker: Lilian Hollard, Université de Reims Champagne-Ardenne, France.
First year Ph.D. student at LICIIS from the University of Reims Champagne Ardenne, development of Artificial Intelligence models for viticulture: Disease identification and yield estimation. Adaptation and optimization of Deep Learning models for Edge and mobile devices.
12:00-12:20
Presentation: A01_4: A Low Power And High-Performance Software Approach to Artificial Intelligence On-Board.
Pablo Ghiglino, Klepsydra Technologies, Switzerland.
Future spacecraft require increased autonomy, more sensors, and better processing power for space exploration, Earth Observation, and robotics. These advancements bring challenges like low determinism, high power consumption, data losses, and response delays. This article presents a novel approach to on-board artificial intelligence (AI) using data pipeline and parallelization techniques to enhance deep neural networks (DNN) in space applications. The results show significantly improved data processing speed and power efficiency, enabling practical AI in space.
Speaker:Pablo Ghiglino, Klepsydra Technologies, Switzerland.
Dr Pablo Ghiglino, Klepsydra founder and CEO. Over 10 years of experience in the finance industry, regarded as a highly skilled software developer. Parallel to his work in finance, Pablo successfully studied for a PhD at the Surrey Space Center in Aerospace Engineering in the field of “High Precision Planetary Landing”. After his studies, he got involved in several aerospace projects, where he seized the opportunity and created Klepsydra to respond to the needs for cutting edge computing by implementing successful software techniques from the highly competitive financial sector in embedded software frameworks for its use in other industries.
Presentation: B01_4: Upstream Mask Tuning of the Plain Vision Transformer for Domain Generalization on Downstream Segmentation Tasks.
Tommie Kerssies, Eindhoven Technical Univerity, the Netherlands.
Deploying perception models on edge devices requires robustness to real-world distribution shifts. We reveal that a fine-tuned plain vision transformer excels in out-of-distribution generalization but is hindered by single-scale features and image-level, as opposed to pixel-level, pretraining. To mitigate this, we employ a mask-tuning step on a large, high-resolution dataset consisting solely of masks, guided by a frozen teacher for semantic context. Our work aims to redirect the domain generalization field towards optimizing pretraining rather than complicating fine-tuning.
Speaker: Tommie Kerssies, Eindhoven Technical Univerity, the Netherlands.
Tommie Kerssies is a PhD candidate in the Mobile Perception Systems Lab at Eindhoven University of Technology, under the guidance of Gijs Dubbelman. Specializing in the generalization capabilities of foundation models, Tommie brings a strong academic background with a master’s in data science and a bachelor’s in computer science.
12:30-13:30
Networking Lunch
13:30-15:30
Session C01: RISC-V-based Edge AI Hardware Acceleration.
Session Co-chairs:
Konstantinos Georgopoulos, Iakovos Mavroidis, Technical University of Crete, Greece
RISC-V-based edge devices are rapidly expanding their scope and application range. They are efficient in executing a variety of AI algorithms close to the application environment, ensuring energy efficiency while maintaining high performance. Furthermore, the open-source architecture of RISC-V enables researchers to customize and enhance it, incorporating low-power AI and security accelerators tailored to the specific needs of edge applications. This session will focus on RISC-V-based edge architectures and their diverse applications.”
Keywords: RISC-V, hardware accelerator, reconfigurable hardware, AI accelerator, HW/SW co-design, energy consumption, HW-based security, ΙοΤ.
Session D01: Edge AI Accelerators Architectures.
Session Co-chairs:
Mario Diaz Nava, STMicroelectronics, France, María Loreto Mateu Sáez, Fraunhofer/IIS, Germany.
Efficient Edge AI system solutions (with highly energy efficient, low latency and high integration) require new HW Processors that support AI advanced features based on neuromorphic architectures implemented in different embedded non-volatile memory and SRAM technologies. This session presents a selection of different AI accelerators capable of addressing different areas of Edge applications. These hardware processors are essential for optimal Edge solutions, but a trade-off between cost and technological features is necessary.
Keywords: HW accelerators, eNVM, FPGA, architecture, neural networks, edge IoT devices, framework, toolchain, dataflow, flexibility, scalability.
13:30-13:50
Presentation: C01_1: How to Accelerate Your Application on RISC-V Using a Coprocessor Interface.
Dr. Christian Herber, NXP, Germany.
RISC-V is becoming the most popular ISA for advanced applications like AI due to its modularity and customizability. However, customization comes with several downsides, like ecosystem fragmentation and design complexity. A clearly defined coprocessor interface can reduce design efforts and improve interoperability. This presentation outlines the use-cases and challenges in a coprocessor interface based on concrete interface definitions.
Speaker: Dr. Christian Herber, NXP, Germany.
Christian Herber the European Principal RISC-V Architect at NXP. Based in Hamburg, his work focuses on innovation management and advanced IP for RISC-V processors in Europe. He has a background in automotive processing, networking, and software. He holds a PhD in Electrical Engineering and an M.Sc. in Electrical Engineering and Information Technology, all from Technical University of Munich, Germany.
Presentation: D01_1: Efficient Mapping of Neural Networks on FPGA With Generated VHDL code.
Thierry Delafontaine, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
FPGAs are an ideal platform to accelerate AI processing at the edge. However, straightforward toolchains to deploy AI accelerators without in-depth knowledge of FPGA programming are lacking. We present a framework that allows the design of custom AI accelerators for FPGAs by generating application specific VHDL code. The framework proposes a dataflow-style hardware mapping that is specific to the NN and optimally uses the available resources. We focused on providing an easy-to-use interface for users and an expandable core that allows developers to add custom operators, optimizers, objectives, metrics, and callbacks.
Speaker: Thierry Delafontaine, Institute of Embedded Systems (InES), ZHAW School of Engineering, Switzerland.
Thierry Delafontaine is a researcher at the Institute of Embedded Systems at the University of Applied Sciences in Zürich. His research topics include embedded AI and accelerating Neural Networks on embedded devices. Currently, he is building the iAIa framework, that he presents at the EEAI conference.
13:50-14:10
Presentation: C01_2: The Beauty of RISC-V Customization for Edge AI HW Acceleration.
Arturo Prieto, Lund University, Sweden.
Edge AI is becoming an increasingly popular topic due to the continuous development of AI applications avoiding the dependencies on-the-cloud. The use of sensitive data or latency constraints push for processing on edge devices. Battery-limited and area-constrained devices offer the opportunity for hardware designers to find efficient architecture solutions. In this presentation, a RISC-V platform is presented as the conductor for the customized integration of hardware accelerators.
Speaker: Arturo Prieto, Lund University, Sweden.
Arturo Prieto received his B. Sc. degree in Electronics Engineering in 2017 from the Polytechnic University of Valencia, Spain, and the M. Sc. degree in Electronic Design in 2019 from Lund University, Sweden. He is currently a PhD student in the Digital ASIC group at the department of Electrical and Information Technology (EIT), Lund University. His doctoral studies focus on beyond-von Neumann architectures for edge computing.
Presentation: D01_2: A Scalable and Flexible Interconnect-based Dataflow Architecture for Edge AI Inference.
Hana KRICHENE. CEA-List, France.
The scalable approach of edge AI inference has increased its computational complexity due to the involvement of multiple properties. This work proposes an interconnect-based dataflow architecture based on many Processing Elements to overcome such problems. The proposed architecture can efficiently handle different convolutions featuring different shapes of input image/feature maps and filters, with data reuse and communication-computation overlap. It is scalable and configurable to adapt to different CNN layers.
Speaker: Hana KRICHENE. CEA-List, France.
Dr. Hana Krichene obtained her Ph.D. Degree in Computer Science in 2015 from the University of Lille. She joined CEA LIST (Saclay, France) in 2017 to participate in the development of the Future E/E Architecture for Automotive Computing Environment. Since 2020, she conducts study research on NoC optimized for AI hardware accelerators of future systems used by the industrial partners of the project. Her current research focuses on developing an advanced dataflow architecture and an environment for mapping and estimating the cost of running a DNN application on a given dataflow architecture.
14:10-14:30
Presentation: C01_3: Heterogeneous SoA Processing on the Edge.
Yannis Papaefstathiou, Aristotle University of Thessaloniki, Greece.
The REBECCA project is developing efficient edge-AI systems that can overcome physical limitations of edge devices and meet regulations and constraints in numerous distinct application domains. This involves designing and developing hardware, software, and middleware that can accelerate computation-intensive parts of both AI and conventional applications, ensure deterministic response times, and satisfy safety constraints. The project is working towards enhancing European strategic autonomy and sovereignty by using an open, royalty-free CPU architecture (i.e., RISC-V), and address performance safety and security concerns associated with IoT and edge devices.
Speaker: Yannis Papaefstathiou, Aristotle University of Thessaloniki, Greece.
Yannis Papaefstathiou is a Professor at ECE School and Aristotle University of Thessaloniki and a co-founder and CEO of Exascale Performance Systems (EXAPSYS). His main research interests are in the design and implementation methodologies for CPS with tightly coupled design parameters and highly constrained resources as well as in heterogeneous High-Performance Computing (HPC) systems and the associated programming/development tools. He was granted a PhD in computer science at the University of Cambridge in 2001, an M.Sc. from Harvard University in 1997 and a B.Sc. from the University of Crete in 1996. He has published more than 100 papers in IEEE and ACM-sponsored journals and conferences. He has participated in more than 15 competitively funded research projects in Europe with a cumulative budget share of more than €9 million.
Presentation: D01_3: DeepEdgeSoC: End-to-end Deep Learning Framework for Edge IoT Devices.
Mhd Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
In the IoT domain, FPGAs are considered a powerful alternative to general purpose graphics processing units (GPGPUs) since their flexible architecture can run the DNNs with much less energy. The enormous amount of effort and time required for the entire end-to-end edge-aware deployment urged us to develop DeepEdgeSoc, an integrated framework for DL design and acceleration. DeepEdgeSoc is an overarching framework under which DNNs can be built. DeepGUI, a visual drag-and-drop DNN design environment, plays an important role in accelerating the network design phase. In DeepEdgeSoc, the networks can be quantized and compressed to suite the underlying edge devices in terms of size and energy.
Speaker: Mhd Rashed Al Koutayni, DFKI Kaiserslautern, Germany.
After receiving his master degree in Embedded Computing Systems at the Technical University of Kaiserlslautern-Landau (RPTU), Mr. Mhd Rashed Al Koutayni joined the Augmented Vision department of the German Research Center for Artificial Intelligence (DFKI) in Kaiserslautern (Germany) as a scientific researcher while pursuing at the RPTU university as a PhD student. His main research focus is FPGA-Based Hardware Acceleration of Deep Neural Networks for Computer Vision.
14:30-14:50
Presentation: C01_4: Why Running Kubernetes on RISC-V is Important for AI Applications.
Antony Chazapis, FORTH, Greece.
Kubernetes is the leading open-source container orchestration platform, widely adopted for easily developing and running scalable applications from the Cloud to the Edge. And now, through K3s (a CNCF-certified distribution), one can also run Kubernetes on devices with processors that use the open RISC-V instruction set. In this talk we will present the process of supporting the new architecture and discuss the prospects this brings for AI applications in future Edge platforms.
Speaker: Antony Chazapis, FORTH, Greece.
Antony Chazapis is a postdoctoral researcher at the CARV Lab of FORTH-ICS. He works on distributed, large-scale platforms for both compute and storage that span the whole computing continuum, from the Cloud to the Edge. His experience and interests include acceleration technologies for Machine Learning/AI in both HPC and the Cloud, Cloud-native software architectures, bridging Cloud with HPC, as well as DevOps and MLOps practices.
Presentation: D01_4: A FeFET-based Non-Volatile Memory and AI Accelerator Macro.
Marko Noack, Ferroelectric Memory GmbH, Germany.
The presentation focuses on the design of a highly configurable AI accelerator macro for convolutional neural networks using ferroelectric NVM as weight storage and which serves as a building block for edge AI systems. We furthermore present a test setup in which the digital part of the circuit was implemented on an FPGA using weights stored on a separate FeFET NVM chip to demonstrate image processing tasks.
Speaker: Marko Noack, Ferroelectric Memory GmbH, Germany.
Marko Noack received the M.Sc. degree in electrical engineering from Technische Universität Dresden, Germany in 2011. He was working on VLSI implementations of neuromorphic circuits before he co-founded the Ferroelectric Memory GmbH (FMC) in 2016, where he is currently Member of Technical Staff and project leader. He works on non-volatile memory circuit designs based on ferroelectrics.
15:00-15:30
Afternoon Coffee/Tea Break
15:30-17:00
Session E01: Trustworthy Edge AI.
Session Co-chairs:
Patrick Pype, NXP Semiconductors, Belgium, Kay Bierzynski, Infineon Technologies AG, Germany.
Trustworthy edge AI addresses the development and deployment of AI systems at the edge of a network with a focus on ensuring system dependability e.g., reliability, security, privacy, and ethical behaviour among other properties. Ensuring trustworthiness in edge AI-based systems requires a multidisciplinary approach that combines expertise in electronics, sensing, IoT, AI, cybersecurity, privacy, ethics, and regulatory compliance. The session addresses the concepts for edge AI trustworthiness and the challenges for developing and deploying edge AI trustworthy systems.
Keywords: explainable edge AI, interpretable edge AI, transparency, edge AI bias mitigation, consent mechanisms, anonymisation, adverbial defence, accountability, verification, validation and testing.
Session F01: Optimisation Methods and Techniques for Neural Networks.
Session Co-chairs:
Gerardo Daalderop, NXP Semiconductors, the Netherlands, Alberto Faro, DEEPSENSING SRL, Italy.
Optimisation methods and techniques for neural networks depend on the specific problem, architecture, and available AI-based HW resources. Experimentation and tuning are required to find the optimal combination for a given neural network task. The session covers various optimisation methods and techniques used in training neural networks to find the set of model parameters that minimise a predefined loss functions.
Keywords: Bayesian neural networks, pruning techniques, energy, size, accuracy, hierarchical-controlling system, neural network processing.
15:30-15:50
Presentation: E01_1: Efficient AI-based Attack Detection Methods for Sensitive Edge Devices and Systems.
Celine Thermann, University of Lübeck, Germany.
Sensitive data is stored on an increasingly large number of edge devices, representing a lucrative target for attackers. Hardware and performance overhead, as well as environmental stresses, play a crucial role in the selection of appropriate solutions. In this research contribution, we are examining the requirements for Artificial intelligence (AI)-based mechanisms and the sensing capabilities of state-of-the-art secure devices.
Speaker: Celine Thermann, University of Lübeck, Germany.
Celine Thermann is a research assistant at the Institute of Computer Engineering at the University of Lübeck, Germany. She received both a Bachelor and Master degree in 2020 and 2022 in IT-Security from the University of Lübeck, respectively. Her main research interest is deep learning and embedded system security, and she focuses on the trustworthiness of AI accelerators.
Presentation: F01_1: Memristor Based Bayesian IMC.
Tifenn HIRTZLIN, CEA-LETI, France.
This presentation examines three methods employing resistive memory’s natural randomness for Bayesian near-memory and in-memory computing. The approaches encompass Bayesian machines optimized for near-memory computing, Bayesian neural networks that capitalize on synaptic randomness, and Bayesian learning using the Metropolis-Hastings Markov Chain Monte Carlo technique. These techniques yield competitive accuracy compared to traditional software methods while enabling decision uncertainty assessment.
Speaker: Tifenn HIRTZLIN, CEA-LETI, France.
Tifenn HIRTZLIN earned a Ph.D. in Nanosciences and Electrical Engineering from the University of Paris-Saclay, France, in 2020, and is currently a Research Engineer at CEA Leti. He is at the forefront of designing intelligent memories for the Internet of Things (IoT), drawing inspiration from bio-inspired concepts such as the Bayesian approach to brain function. Tifenn’s primary responsibility involves designing electrical circuits for application-specific integrated circuits (ASICs) that exploit non-volatile memory devices, and evaluating their performance, contributing significantly to the advancement of IoT technology.
15:50-16:10
Presentation: E01_2: Explainability and Interpretability Concepts for Edge AI Systems.
Angelo Genovese, Università degli Studi di Milano, Italy.
The presentation gives an overview of the recent techniques for AI explainability and interpretability, with a specific focus on Edge computing. The presentation will discuss the motivations for explainability and interpretability, a taxonomy of the latest methodologies, the most common benchmarking approaches, as well as challenges, open issues, and research trends.
Speaker: Angelo Genovese, Università degli Studi di Milano, Italy.
Angelo Genovese received the Ph.D. degree in computer Science from the Università degli Studi di Milano, Italy, in 2014. He has been an Associate Professor in Computer Science with the Università degli Studi di Milano since 2022. He has been a Visiting Researcher with the University of Toronto, Toronto, ON, Canada. Original results have been published in over 60 papers in international journals, proceedings of international conferences, books, and book chapters. His current research interests include signal and image processing, artificial intelligence, industrial and environmental monitoring systems.
Presentation: F01_2: Trade-Off Analysis of Pruning Methods for Compact Neural Networks on Embedded Devices.
Duc V Le, University of Twente, The Netherlands.
Pruning of neural networks is a technique often used to reduce the size of a machine learning model, as well as to reduce the computation cost for model inference. This study analyses four pruning techniques for efficient reduction of machine learning model size, based on compression-accuracy balance. Evaluation includes impact on energy use during inference on a Raspberry Pi 4B using MobileNetV2. We examine trade-offs among energy, size, and accuracy, aiding engineers in selecting optimal pruning for embedded ML models. The research is expected to provide engineers with a reference providing guidance upon deciding what pruning technique to use for a machine learning model to be deployed on an embedded device.
Speaker: Duc V Le, University of Twente, The Netherlands.
Dr. Le Viet Duc is Assistant Professor at the Department of Computer Science at the University of Twente. Inspired by the way the human brain learns the world, he is eager to develop sustainable AI models for that is robust to change in data distribution caused by new environments, bridging the gap between lab and word. His focus is on finding efficient methods to learn, transfer and distil knowledge hidden in different datasets, especially unlabelled ones. As he has 10 years R&D experience in developing smart and wearable and industrial devices, he is also interested in applying his research to industrial products for a sustainable world.
16:10-16:30
Presentation: E01_3: Approximate Computing in a Federated Learning Context: Gains Beyond Energy.
Ihsen Alouani, Université Polytechnique Hauts-De-France, France.
In energy-constrained systems at the edge, the adoption of federated learning has garnered significant attention due to its dual capacity to address privacy concerns while achieving accurate model training without resource-intensive methodologies. The Approximate Computing paradigm, characterized by leveraging controlled imprecision to achieve computational efficiency, has emerged as one of the solutions towards a trade-off between performance and energy consumption in fault-tolerant applications. Expanding on these foundational concepts, the presentation proposes an innovative approach, i.e., “Approximate Federated Learning.” Within this framework, nodes collaboratively undertake training and inference procedures utilizing approximate computing techniques, thereby accentuating energy efficiency gains. The study extends the benefits beyond energy conservation, highlighting substantial security enhancements achievable within this amalgamation. This holistic framework advances energy optimization in edge scenarios and bolsters the overall security and trustworthiness of edge AI systems.
Speaker: Ihsen Alouani, Université Polytechnique Hauts-De-France, France.
Ihsen Alouani received his master’s degree from the National Engineering School of Sousse, Tunisia, and his Ph.D. in 2016 from the Université Polytechnique Hauts-de-France. After his PhD, he was a Postdoc at Technical University of Munich, Germany followed by an Associate Professor at IEMN Laboratory (CNRS-8520), Université Polytechnique Hauts-de-France in 2018. He served as Guest Editor for several journals including Microprocessors & Microsystems and is a member of the Program Committee of several international conferences including Design Automation Conference (DAC) as well as Asia and South Pacific Design Automation Conference (ASP-DAC). His research interest includes machine learning security and privacy, Approximate Computing, and Embedded Systems reliability and security.
Presentation: F01_3: SENECA: Flexible and Scalable Neuromorphic Processor for Efficient Event-driven Neural Processing.
Guangzhi Tang, imec, the Netherlands.
The lack of flexibility in most neuromorphic architecture designs results in significant performance loss and inefficient memory usage when mapping various neural network algorithms. We present SENECA, a digital neuromorphic architecture that balances the trade-offs between flexibility and efficiency using a hierarchical-controlling system. A SENECA core contains two controllers, a flexible RISC-V-based controller, and an optimized controller (Loop Buffer). This flexible computational pipeline allows for deploying efficient mapping for various neural networks, on-device learning, and pre-post processing algorithms. The hierarchical-controlling system introduced in SENECA makes it one of the most efficient neuromorphic processors for event-driven neural network processing.
Speaker: Guangzhi Tang, imec, the Netherlands.
Guangzhi Tang is a researcher at imec based in Eindhoven, the Netherlands. His research focuses on neuromorphic computing, hardware-aware neural networks, and robotics. He has extensive experience working in the field of neuromorphic computing to explore the real-world advantages of digital neuromorphic architectures. He received his Ph.D. and M.Sc. degrees in Computer Science from Rutgers, the State University of New Jersey in the United States, and his B.Sc. degree in Computer Science from Nanjing University in China.
16:30-16:50
Presentation: E01_4: Shaping Edge AI That Flows With Humans: Designing Trustworthy and Useful AI Applications.
Peter Moertl, Virtual Vehicle Research GmbH, Austria.
The introduction of artificial intelligence solutions in industrial work processes leads in most cases to changes in human tasks and responsibilities rather than complete autonomous operations. However, technology-centered AI introductions that are currently still common, frequently overlook the role of the human operator and the involved organizations, resulting in significant AI effectiveness reductions and even social backlashes, slowing down progress, AI uptake, and real benefit-accrual. In the InSecTT project we have investigated various use cases for AI applications in different domains including medical, transportation, IT, and manufacturing and have formulated a framework for successful AI introductions that centers around human oversight. In this presentation we will present this framework and give examples for its application in multiple domains.
Speaker: Peter Moertl, Virtual Vehicle Research GmbH, Austria.
Dr Peter Mörtl is the manager of the human the factors teams at Virtual Vehicle Research center in Graz, Austria. His focus is on preparing and facilitating human and system integration for technological developments in automotive and AI applications, specifically focusing trust, user modeling, and user acceptance by applying human factors and systems engineering. Previously he has worked as systems engineer and human factors researcher in aviation projects in the United States to develop flight-deck and air traffic control applications. He is also Lector for Human-Systems Integration at the University of Graz, Austria.
Presentation: F01_4: Energy Efficient Deep-Edge Computing through Hardware Machine Learning.
Marcello Coppola, ST Microelectronics, France.
It is currently challenging to accomplish high data processing with low latency, because large-scale, persistent, high-bandwidth connection is often required to gather huge volumes of data from platforms and sensors operating at the edge to allow processing in the cloud. Internet-of-Things (IoT) devices deployment in remote locations for environmental monitoring with machine learning (ML) techniques, Internet-of-Medical Things, IoT-based precision agriculture, all can operate with limited battery, while robustness and reliability is required.
Speaker: Marcello Coppola, ST Microelectronics, France.
Marcello Coppola is a Technical Director at STMicroelectronics, in charge of the implementation of several research projects, including overseeing day-to-day field-based activities related to System-On-Chip Consumer Space, Automotive, IoT, Healthcare, with particular emphasis on Network-on-Chip, heterogenous multi-core hardware and software architectures which include CPU, GPU and accelerators.
17:00
End
20:00
Dinner
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
| 11:00-12:30 |
| 11:00-11:20 |
| 11:20-11:40 |
| 11:40-12:00 |
| 12:00-12:20 |
| 12:30-13:30 |
| 13:30-15:30 |
| 13:30-13:50 |
| 13:50-14:10 |
| 14:10-14:30 |
| 14:30-14:50 |
| 15:30-16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 16:00 |
| 17:30 |
| 20:00 |
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation A01_1 - Name Presentert, Company Affiliation, Country
Presentation A01_2 - Name Presentert, Company Affiliation, Country
Presentation A01_3 - Name Presentert, Company Affiliation, Country
Presentation A01_4 - Name Presentert, Company Affiliation, Country
| Networking Lunch |
Parallel Session C01: Title - Room C01
Presentation C01_1 - Name Presentert, Company Affiliation, Country
Presentation C01_2 - Name Presentert, Company Affiliation, Country
Presentation C01_3 - Name Presentert, Company Affiliation, Country
Presentation C01_4 - Name Presentert, Company Affiliation, Country
| Afternoon Coffee/Tea Break |
Parallel Session E01: Title - Room E01
Presentation E01_1 - Name Presentert, Company Affiliation, Country
Presentation E01_2 - Name Presentert, Company Affiliation, Country
Presentation E01_3 - Name Presentert, Company Affiliation, Country
Presentation E01_4 - Name Presentert, Company Affiliation, Country
| End |
| Dinner |
Parallel Session B01: Title - Room B01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Presentation B01_1 - Name Presentert, Company Affiliation, Country
Presentation B01_2 - Name Presentert, Company Affiliation, Country
Presentation B01_3 - Name Presentert, Company Affiliation, Country
Presentation B01_4 - Name Presentert, Company Affiliation, Country
Parallel Session D01: Title - Room D01
Presentation D01_1 - Name Presentert, Company Affiliation, Country
Presentation D01_2 - Name Presentert, Company Affiliation, Country
Presentation D01_3 - Name Presentert, Company Affiliation, Country
Presentation D01_4 - Name Presentert, Company Affiliation, Country
Parallel Session F01: Title - Room F01
Presentation F01_1 - Name Presentert, Company Affiliation, Country
Presentation F01_2 - Name Presentert, Company Affiliation, Country
Presentation F01_3 - Name Presentert, Company Affiliation, Country
Presentation F01_4 - Name Presentert, Company Affiliation, Country
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
8:30-9:00
Welcome and Registration
9:00-10:30
Plenary Session III: Edge AI Applications
9:00-09:30
Convergence of Technologies Patrick Pype, NXP Semiconductors, Belgium
In the advent of an intelligent and connected world, the underlying, enabling technologies have made headline news. Modern societies’ dependence on semiconductor technologies, microprocessors, and system-on-chip (SoC) has become painfully obvious in a very short time. In parallel, the evolved digital infrastructure and related technologies are becoming undisputable cornerstones for economic development. Autonomous driving, electrification of vehicles, SW defined vehicles, internet-of-everything (IoX), artificial intelligence and edge computing, new communication paradigms, use of open-source as e.g. RISC-V, are all elements that lead to a transformation of the semiconductor towards system integration landscape. It is important to see how this all will converge towards solutions leading towards a better life on this earth.
Presenter: Patrick Pype, NXP Semiconductors, Belgium.
Patrick is Director of Strategic Partnerships at NXP Semiconductors and has held this position since 2012. He began his career at IMEC in Belgium in 1986. In 1996 he started the start-up company CoWare on hardware/software co-design which was later acquired by Synopsys. Thereafter he worked at Philips in different positions. Patrick is Chairman of the AENEAS Technical Expert Group since 2018 and he is member of the INSIDE Steering Board since 2015. He was co-chair of the overall ECSEL Strategic Research Agenda 2019-2021 and is also co-chair of the chapter on “Transport & Mobility”. Patrick was Chairman of the Engineering Alumni Association of the KULeuven from 2005-2008. Patrick has contributed several papers in two books published by Springer on “Automated Driving’’ and “Internet of Vehicles’’. He is author of the book “Zest for Opera – Unleash your Leadership”, which was published in October 2018. The book offers inspiring insights from the world of opera for leaders in business & politics, coaches, teachers and students.
9:30-10:15
Smart sensors: paving the way for a sustainable future. Lisa Trollo, STMicroelectronics, Italy
Sensors play a crucial role in connecting the physical and digital worlds, facilitating many of our interactions with the environment. As technology advances, people expect it to prioritize human-centeredness and foster creativity without being intrusive. To achieve this, sensors are being designed with artificial intelligence to optimize edge computing, making them a vital component in the journey towards a sustainable future. These smarter sensors can create ultralow power systems that offer innovative solutions to environmental and social challenges.
Presenter: Lisa Trollo, STMicroelectronics, Italy
After joining ST in 2012, Lisa contributed to several success stories in consumer, automotive and industrial markets, such as the massive Inertial Measurement Unit (IMU) product deployment in the smartphone industry. Always working to help developers find the best solution for their applications, Lisa is currently in charge of our MEMS Ecosystem as well as Marketing Communication where she plays a key role in defining ST’s Artificial Intelligence strategy for MEMS and sensors. Lisa graduated with a Master’s degree in Telecommunication Engineering in 2006 from The University of Trento, Italy. technologies.
10:15-10:30
Projects Pitch Presentations
10:30-11:00
Morning Coffee/Tea Break
11:00-17:00
Parallel Sessions
11:00-12:30
Session A03: Intelligent Mobility Systems.
Session Co-chairs:
Christian Ohms, Mercedes-Benz AG, Germany, Alfred Höß, OTH Amberg-Weiden, Germany.
The integration of AI capabilities at the edge of a network or within vehicles, transportation, and mobility systems to enable autonomous mobility to enhance the capabilities and increase safety and reliability. These systems leverage AI algorithms and real-time data processing to make decisions, control vehicles, and navigate autonomously without relying heavily on remote cloud-based computation. The session includes presentations that address key aspects and applications of autonomous edge AI mobility systems.
Keywords: autonomous vehicles, edge AI for perception, path planning, localisation, advanced driver assistance systems, V2X communication.
Session B03: Digital Industry.
Session Co-chairs:
Falk Hoffmann, NXP Semiconductors, Germany, Claus Lenz, Cognition Factory GmbH, Germany.
The transformation and integration of digital technologies, such as data analytics, automation, the IoT, AI, and immersive technologies require the deployment of these technologies into various sectors of the economy and industrial processes. The session presents various use case employing edge AI technologies in industrial manufacturing processes that drive technological advancements and changing manufacturing patterns with a profound impact on the economy, markets, and society.
Keywords: advanced AI-based manufacturing, maintenance, repair, and operations, functional safety edge AI systems, intelligent radar signal processing.
11:00-11:20
Presentation: A03_1: AI for Connected and Shared Mobility.
Jochen Koszescha, Infineon Technologies AG, Munich, Germany.
The presentation is giving an overview of AI4CSM project that focuses on the design of highly optimized and automated components for digital mobility that allows smart and energy efficient fleet operations in terms of smart routing and platooning achieving low latency in the vehicle intelligence and perception systems. This means not only smart sensor fusion, but also reducing the level of complexity by using functional architectures and trustable automotive AI, to foster Europe’s competitiveness in the global automotive market. The vision of AI4CSM is to build Europe’s intelligent electronic component and systems for ECAS 2030 vehicles supporting European mass market production, manufacturability and scalability based on the Green Deal principles (incl. Vision Zero and Safe System) to address the sustainable urbanization challenge defined by the United Nations.
Speaker: Jochen Koszescha, Infineon Technologies AG, Munich, Germany.
Jochen Koszescha is a senior director at Infineon Technologies AG with over 30 years of experience in Electronic Components and Systems at the Aviation and Transport Industries. He has been coordinating several research and development projects at Infineon.
Presentation: B03_1: Distributed Multisensory System, from Pre-Development to Production.
Claus Lenz, Cognition Factory GmbH, Germany.
This talk showcases the industrial use-case of a distributed multisensory system applied in the field of Maintenance, Repair, and Operations. It highlights the advantages and challenges of a distributed architecture of sensor nodes, capable of independent functioning and data processing. The presentation covers the development cycle from the Requirement formulation phase, up to Deployment and Maintenance. Emphasis is made on certifiable approach to AI-based solutions, and delegation of intelligence to the edge, as well as an architectural design that is scalable both vertically and horizontally.
Speaker: Claus Lenz, Cognition Factory GmbH, Germany.
CEO and co-founder of Cognition Factory GmbH. PhD in human-robot collaboration at Technische Universität München (TUM). Experienced within research projects aimed at digitising the industry: AI4DI, Providentia++, DiProLea.
11:20-11:40
Presentation: A03_2: Natural Language Conditioned Planning of Complex Robotics Tasks.
Janis Arents, Institute of Electronics and Computer Science, Riga, Latvia.
As natural language processing advances in the field of robotics, enabling seamless human-robot interaction, it becomes imperative to identify the most effective approach for conditioning complex robotics tasks using natural language commands. This chapter reviews various state-of-the-art methods for natural language-conditioned planning, with a particular focus on mobile manipulation. We explore and review different architectures and techniques to comprehend, interpret, and execute natural language commands. Challenges are identified along the way, and conceptual architecture is proposed to tackle them in an efficient manner.
Speaker: Janis Arents, Institute of Electronics and Computer Science, Riga, Latvia.
Researcher Janis Arents has been working in EDI since 2016. Graduated Riga Technical University (Riga, Latvia) in the year 2018 with professional master’s degree in electrical engineering. Currently studying in Riga Technical University in Computerised Control of Electrical Technologies doctors’ degree program. Main research topics are connected with innovative digital technology usage in industrial applications, including computer vision and artificial intelligence-based systems for industrial process automation.
Presentation: B03_2: Outlook on the Deployment and Execution of Mixed-Criticality Edge-AI Applications.
Suzana Veljanovska, Zurich University of Applied Sciences, Switzerland.
Functional safety-relevant edge AI systems are the cutting of AI research, and the technological gap is significant. In this presentation we approach the issue of creation and deployment of both typical AI applications and functional safety systems. We pay particular attention to the merging of the two workflows and present options and ideas to achieve this.
Speaker: Suzana Veljanovska, Zurich University of Applied Sciences, Switzerland.
Suzana Veljanovska is a research assistant at Zurich University of Applied Sciences (ZHAW) working in the field of dependable AI.
11:40-12:00
Presentation: A03_3: Dynamic LoRa Broadcasting and Obstacle Prediction.
Dimitris Georgiadis, Harokopio University of Athens, Greece.
This presentation covers significant achievements in the development of a LoRa-based emergency communication network for highly automated vehicles, validated through simulations using NS3. The simulation comprises stationary infrastructure, including a gateway a network server, and multiple mobile vehicles acting as end devices. Concluding in this research project we present a significant advancement in LoRa-based emergency communication simulations, with achievements in realistic mobility modelling, communication protocols, and visualization tools. Ongoing efforts are focused on enhancing emergency responsiveness focused on obstacle recognition, optimizing communication reliability, and accelerating the network’s adaptability to dynamic scenarios.
Speaker:Dimitris Georgiadis, Harokopio University of Athens, Greece.
Dimitris Georgiadis is a research associate at Harokopio University of Athens. His works focus on 5G D2D single-cell communications and minimization of energy consumption, UAV swarm communications and currently studying vehicle behavior. Graduate from Harokopio University of Athens from the Department of Informatics and Telematics and is now a postgraduate Student of Master Programme (MSc) MPhil in Computer Science and Informatics.
Presentation: B03_3: RANet: An Autoencoder for High-Resolution Range-Angle Maps in Multistatic Radar.
Chen Liu, Technische Universität Dresden, Germany.
Advanced radar perception tasks, such as target recognition and tracking, often impose strict requirements on radar imaging quality. However, conventional radar imaging methods are constrained by spatial angular resolution, which is limited by the number of virtual channels composed by the antenna array, thus leading to suboptimal performance of low-cost, low-resolution radar systems. This paper introduces an innovative deep learning model: RANet, which leverages the raw sampling signals from multiple radar sensors to generate high-resolution range-angle maps in an end-to-end manner. RANet not only partially replaces the traditional radar processing chain but also demonstrates superior performance in downstream tasks such as object counting and tracking. Experimental results indicate that monitoring up to four indoor individuals achieves an accuracy of 96.4% by exploiting accumulated temporal information of only two 1T3R (one transmitter, three receivers) radar sensors.
Speaker: Chen Liu, Technische Universität Dresden, Germany.
Chen Liu received the BSc degree and MSc degree in electrical engineering in 2012 and 2016, respectively. He is currently working toward the PhD degree in Technische Universität Dresden, Germany. His research interests include intelligent radar signal processing, efficient deep learning on the edge and neuromorphic computing in automotive.
12:00-12:20
Presentation: A03_4: Edge Perception in Autonomous Driving.
Gerhard Dorn, Virtual Vehicle Research GmbH, Austria.
The talk gives a small introduction to our current work on the autonomous driving Robotaxi within the EU project AI4CSM and focusses on development and integration techniques of edge perception algorithms using Julia for machine learning tasks. Challenges regarding data quality, data sources, uncertainties, and run-time issues as well as opportunities regarding fast prototyping and integration will be discussed.
Speaker:Gerhard Dorn, Virtual Vehicle Research GmbH, Austria.
Gerhard Dorn is a senior scientist at Virtual Vehicle Research GmbH with a background in mathematics and computational physics.Within the project AI4CSM Gerhard works on LIDAR perception and the integration into the test vehicle Ford Mondeo.
Presentation: B03_4: Smart Semiconductor Manufacturing Assistant (SSMA) – An Efficient Quality Management Supporting Tool.
Silke Braun, Infineon, Austria.
Smart Semiconductor Manufacturing Assistant (SSMA) as an efficient quality management support tool is an important step in the direction of establishing causal links between production-related data such as production equipment and process parameters, data resulting from process experiments as well as knowledge from FMEAs. This allows potential risks to be quickly identified, predicted, and classified to support the engineer in decision making.
Speaker: Silke Braun, Infineon, Austria.
Silke Braun completed her studies in microsystems engineering at Mittweida University of Applied Sciences (Germany). In 2017, she started working at Infineon Technologies Dresden as a defect density and metrology engineer; since 2018 in Villach (Austria). She strengthened her knowledge in the field of quality management, among others for new technologies like GaN. Since 2022, she focuses her work on the field of funding and works in various EU-funded projects.
12:30-13:30
Networking Lunch
13:30-15:00
Session C03: Autonomous Vehicle Domains Applications.
Session Co-chairs:
Jochen Koszescha, Infineon Technologies AG, Germany, George Dimitrakopoulos, Harokopio University of Athens, Greece.
The integration of AI capabilities at the edge of a network or within vehicles, transportation, and mobility systems to enable autonomous mobility to enhance the capabilities and increase safety and reliability. These systems leverage AI algorithms and real-time data processing to make decisions, control vehicles, and navigate autonomously without relying heavily on remote cloud-based computation. The session includes presentations that address key aspects and applications of autonomous edge AI mobility systems.
Keywords: autonomous vehicles, edge AI for perception, path planning, localisation, advanced driver assistance systems, V2X communication.
Session D03: Digital Society.
Session Co-chairs:
Fetze Pijlman, Signify, the Netherlands, Silke Braun, Infineon, Austria.
The deployment of AI capabilities directly at the edge of the network or on devices within a digitally connected society plays a pivotal role in shaping how technology interacts with individuals and communities, enabling real-time decision-making, enhancing user experiences, and ensuring data privacy. The session focus on several presentations that highlight the use of edge AI in digitising the society by promoting efficiency, responsiveness, and data privacy while enhancing the quality of life and safety for individuals and communities advancing the development of a more connected and intelligent world.
Keywords: smart working environments, digital healthcare, augmented reality, event-based intelligence, environmental monitoring.
13:30-13:50
Presentation: C03_1: SNN-based Accelerators for Low-latency, Ultra-low-power Inference at the Edge.
Maxim Rykunov, imec, Belgium.
State-of-the-art currently focuses on conventional deep neural network architectures for classification lacking in power, performance, and scalability especially for the processing at the edge. In this talk we present two designed and evaluated neuromorphic prototypes to be demonstrated in the AI4CSM project. One of these solutions focuses on a HW/SW co-design based on a spiking neural network and embedded processing for automatic detection, tracking and classification of multiple vulnerable road users. The second one targets the development of a programmable analogue-mixed signal accelerator platform for always-on detection and recognition of time-series patterns in sensor data. Neuromorphic SNN-based approaches are critical in such applications, where we demonstrate novel energy-efficient platforms based for low-latency, ultra-low-power inference at the edge.
Speaker: Maxim Rykunov, imec, Belgium.
Maxim Rykunov received the Ph.D. degree in electrical and electronic engineering from Newcastle University, Newcastle, U.K., in 2013, in asynchronous microprocessors and design of power efficient systems. In 2014, he joined IMEC, Leuven, Belgium, where he is currently a Project Manager in Advanced RF group. His research interests include real-time embedded systems for heterogeneous multisensor data fusion and energy-efficient machine learning for high-resolution radars.
Presentation: D03_1: On the Move Biometrics at the Edge: Security Meets Convenience.
Emanuele Maiorana, Roma Tre University, Italy.
Biometric recognition systems are used to validate the identity of an individual and among the main candidates to be used in contexts with high security requirements (e.g., airports, or border control). Interacting with a biometric recognition system requires a user to stop at a terminal for the time necessary to acquire the relevant traits. In contexts requiring the management of large flows of people, such procedure could lead to the creation of long lines, resulting in discomfort for the involved subjects. To deal with this issue, we propose an “on-the-move” solution where subjects to not need to stop at a gate during the recognition process, since the interested characteristics can be captured while users continue to walk, thus increasing the system throughput and improving passengers’ convenience. The required processing is performed at the edge, thus realising a reliable, low-cost, and convenient system to perform security checks within an airport. In more detail, the proposed system relies on a multimodal approach exploiting face traits to trigger the recognition process at a distance, and subcutaneous vein patterns to guarantee the desired robustness against malicious spoofing attempts.
Speaker: Emanuele Maiorana, Roma Tre University, Italy.
Emanuele Maiorana received the Ph.D. degree in biomedical, electromagnetism, and telecommunication engineering from Roma Tre University, Rome, Italy, in 2009, with European Doctorate Label. He is currently an Assistant Professor with the Department of Industrial, Electronic and Mechanical Engineering, Roma Tre University. His research interests are in digital signal and image processing, with a specific emphasis on biometric recognition. Dr. Maiorana was a recipient of the Lockheed Martin Best Paper Award for the Poster Track at the IEEE Biometric Symposium 2007, the Honeywell Student Best Paper Award at the IEEE Biometrics: Theory, Applications and Systems Conference 2008, and the Best Paper Award at the 11th International Conference on Pattern Recognition Applications and Methods (ICPRAM) 2022. He was the General Chair of the 9th IEEE International Workshop on Biometrics and Forensics (IWBF) 2021. He is an Associate Editor of the IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY.
13:50-14:10
Presentation: C03_2: Robust Propulsion System for Shared Connected Mobility.
Christian Ohms, Mercedes-Benz AG, Germany.
The talk will address developments in drivetrain technology with regard to efficiency, diagnostic methods and charging components as well as current trends in the automotive industry.
Speaker: Christian Ohms, Mercedes-Benz AG, Germany.
Christian Ohms received his diploma in electrical engineering from the University of Applied Sciences in Kiel in 2008. After holding different engineering positions in the automotive industry, he joined the group of power electronics advanced engineering at Daimler AG, now Mercedes-Benz AG, in 2015. His main tasks include component design for future drivetrain concepts and coordination of vehicle demonstrator projects. He has been actively collaborating for various funded projects like HV-ModAL, 3CCar and AutoDrive.
Presentation: D03_2: AI-enabled Resource Optimization in Critical Computing Infrastructures.
Prodromos Makris, National Technical University of Athens, Greece.
The seamless orchestration of resources across the edge-cloud continuum has become a cornerstone in modern critical computing infrastructures. Artificial Intelligence (AI) plays a pivotal role in enabling the fast and efficient management of diverse and demanding workloads. However, maintaining data security, while optimizing resources can be a significant challenge especially in multi-domain environments. AI emerges as a key enabler as it not only increases the robustness and reliability of the distributed computing environments but also contributes in designing resilient orchestration mechanisms across various critical computing domains. The integration of AI within the common resource optimization frameworks, significantly elevates the capabilities of a distributed computing infrastructure and can ensure a future where resource optimization and security are harmoniously balanced to meet the demands of the digital era.
Speaker: Prodromos Makris, National Technical University of Athens, Greece.
Dr. Prodromos Makris is a senior researcher at ICCS/NTUA (National Technical University of Athens), Greece. He holds a 5-year BSc (2007), MSc (2009), PhD (2013) from University of the Aegean, Greece. During the last years, he has been actively participating and leading several national and EC-funded research and innovation projects collaborating with many industrial and academic partners in the EU area. His research interests include energy efficient ICT, resource allocation in IoT-edge-cloud continuum and context-aware mobile/wireless networking in the 5G/6G era. He has more than 60 publications in international conferences and journals. He has recently served as technical manager of FP7 VIMSEN-GA-619547, H2020-GA-731767 SOCIALENERGY and H2020-GA-863876 FLEXGRID projects. He also serves as Adjunct Lecturer at the University of the Aegean.
14:10-14:30
Presentation: C03_3: Reward Shaping for Safe Reinforcement Learning Controllers.
Lorenz Klampfl, AVL List GmbH, Austria.
The performance of controllers based on reinforcement learning heavily depends on designing appropriate reward functions. With poorly designed rewards, an agent can behave unsafely, non-optimal, learn how to obtain the reward without achieving the desired goals, or take too long to learn a policy. In this talk, we give an overview of formal specifications to automatically provide locally shaped rewards. The advantage of such an approach is that we can address all the above-mentioned challenges efficiently.
Speaker: Lorenz Klampfl, AVL List GmbH, Austria.
Lorenz Klampfl holds an M.Sc. in Aviation (2017), is currently pursuing a Ph.D. in Computer Sciences at Graz University of Technology and works as a Simulation Engineer at AVL List GmbH. He is conducting research on quality assurance methodologies for autonomous cyber-physical systems, including the reliability and trustworthiness of neural networks and reasoning techniques for autonomous systems, at the Institute of Software Technology. In his role as a Simulation Engineer, he focuses on the development of ADAS/AD verification and validation methodologies in virtual environments as well as on the edge.
Presentation: D03_3: Quantization- and HW-aware Neural Architecture Search for EdgeAI Solutions.
Gerardo Daalderop, NXP Semiconductors, the Netherlands.
To a significant extent, future smartness in devices of the digital society will originate from neural networks embedded in these devices. However, today typically neural network (NN) models are solely trained to improve the performance of the prediction-the primary task-, without considering the embedded constraints of the devices in which these models will be deployed. Afterwards in a separate process, the models are adapted to fit memory, computation, and performance requirements. The presentation introduces the concept of multi-objective quantization-aware neural architecture search, wherein optimization techniques like quantization and pruning, are combined with HW-aware neural architecture search (NAS). NAS has become a very successful technique for optimizing neural networks, and in HW-aware NAS the properties of the embedded device are brought into the loss-function and training loop, allowing optimization both towards the primary task, as well as secondary objectives being non-functional requirements. A few examples will be given highlighting the benefits of the approach.
Speaker: Gerardo Daalderop, NXP Semiconductors, the Netherlands.
Gerardo Daalderop (PhD, MBA) is a technical director at NXP Semiconductors, having in recent years built-up an AI ML Innovation Center in its Central Technology Office, focusing on AI ML innovations for the Edge beyond next generation products, and across segments as mobility, industrial, home/office, IoT. Before this, he has set up and led groups, roadmaps, programs, and projects in the areas of highly automated driving and safe mobility, safe and secure wireless, telematics and multi-standard communication solutions, and the connected consumer, always with the guiding principle to identify and address highly valued use-cases and user-needs with differentiating technologies, tested and demonstrated in practical trial-environments. As such he has also participated in numerous national and transnational cooperation projects.
14:30-14:50
Presentation: C03_4: Connectivity and Cognitive Communication for Future Shared Mobility.
Gerald Fritz-Mayer, TTTech Auto AG, Austria.
The talk presents the current developments w.r.t. cognitive communication which enables AI-powered methods to access data from both the edge and the cloud, providing fast, reliable, and low-latency data connections. The presentation will showcase the respective demonstrators of (a) proof of concept communication platform and (b) novel technologies to prove wireless data transmission on 28GHz mmW frequency.
Speaker: Gerald Fritz-Mayer, TTTech Auto AG, Austria.
Gerald Fritz-Mayer (male) graduated in computer science at the Technical University of Graz and has an MBA degree on Entrepreneurship and Innovation from the Vienna University of Economics and Business. He joined TTTech as senior coordinator R&D project 2017 and is concerned with autonomous system in the field of manufacturing and autonomous driving. Beside that he participated in several national and European research projects, he was coordinating the innovation in various national research projects in the field of assistive system for future manufacturing systems. Gerald Fritz-Mayer is an IEEE member and serves in different officer roles in the robotic chapter Austria (RAS) and the Technology and Engineering Management Society (TEMS) Austria.
Presentation: D03_4: Accelerating Magnetic Resonance Imaging on Heterogeneous Edge Compute Platforms.
Jan Langer, Fraunhofer ENAS, Germany.
Simulating magnetic resonance imaging (MRI) is a computationally expensive task, that is based on solving the Bloch equations for every voxel and time step. Ideally, the simulation runs in real time on premise with confidential customer data, hence requiring a powerful edge acceleration platform. We developed a fully software-programmable and scalable solution for FPGA platforms achieving multiple magnitudes of performance increase. In this talk we will illustrate the opportunities and challenges we faced during this process.
Speaker: Jan Langer, Fraunhofer ENAS, Germany.
Jan Langer (MS’05 and PhD’11 in Electrical Engineering) leads the artificial intelligence team at Fraunhofer ENAS. Before that, he worked as a senior researcher and processor architect at Xilinx in the heterogeneous multi-core architectures team which developed the AI Engine. His current research at ENAS includes creating hybrid digital twins for semiconductor processes, as well as accelerating algorithms on edge AI platforms.
15:00-15:30
Afternoon Coffee/Tea Break
15:30-17:00
Edge AI Cluster Internal Meeting
17:00
End
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.
Performance models and energy-optimal scheduling of DNNs on many-core hardware with dynamic power management
Bernhard Vogginger, Florian Kelber, Shambhavi Balamuthu Sampath, Johannes Partzsch, Christian Mayr
13.10.2022
Workshop on Compilers, Deployment, and Tooling for Edge AI
Parallel Session A01: Title - Room A01
This presentation highlights on the importance of properly architecting the AI workflows for the design, development, and deployment of predictive maintenance applications in industrial applications at the edge. It is advocated that the selection of the frameworks/platforms employed for the task largely depend on the application, IIoT devices and their physical operating environments, rather than on the available knowledge and experience of the people, as suggested in similar literature. For this purpose, it was employed several existing AI frameworks and inference engines, that permit end-to-end solutions with various degree of automation and integration with Arm® Cortex®-M-based MCUs. The presentation addresses the verification and validation aspects of model design, development, and deployment of AI-based industrial applications and provides quantitative and qualitative insights that can ultimately allow the use of the right framework. The use case presented is a classification for predictive maintenance based on the vibration of generic rotating equipment (such as motors with vibration pumps, fans, compressors), common to many industrial applications, such as manufacturing.
Speaker
Ovidiu Vermesan (SINTEF, Norway)
Ovidiu Vermesan is Chief Scientist at SINTEF Digital, Oslo, where he is involved in applied research on future edge autonomous intelligent systems and edge AI, wireless sensing devices and networks, smart systems integration, microelectronics design of integrated systems (analogue and mixed signal), IIoT. He holds a PhD in Microelectronics and a Master of International Business (MIB). His applied research activities focus on advancing edge AI processing, embedded electronics, wireless and smart sensing technologies, and the convergence of these technologies in different industrial sectors.