Università degli Studi di Milano (UMIL) is researching cutting-edge visual AI models on tiny, low-powered devices. As industries increasingly demand efficient and explainable AI solutions, UMIL is tackling the critical challenges of optimized performance, explainability, and minimized training time.
Visual AI model for industrial applications: Automatically inferring the quality of industrial samples
Here’s how UMIL is driving innovation in this space:
- Training optimization: UMIL has proposed novel techniques that reduce the size of training datasets, enabling the training of visual AI models in just minutes. Despite this dramatically reduced training time, these models achieve performance comparable to those trained over several hours or even days.
- Enhancing explainability: To address the need for transparent AI, UMIL is proposing methods that make neural networks more explainable and interpretable. These methods highlight the key areas responsible for decision-making processes, improving trust and usability in critical applications. In this regard, UMIL’s research is mainly on disentangling objects from their backgrounds. This novel disentanglement approach not only improves explainability but also enhances model performance on specific tasks.
Features disentanglement improves visual explainability.
These advancements are proving especially valuable in industrial applications where low-powered devices are essential. A key example is the production industry, where AI models must analyze images coming from the cameras efficiently and reliably without consuming significant computational resources.
Key benefits of UMIL’s approach:
- Reduced training time: Models can be prepared in minutes, enabling faster iteration and deployment.
- Low-powered device compatibility: Ideal for industries requiring energy-efficient solutions.
- Improved explainability: Transparent decision-making processes build confidence in AI solutions.
Blog signed by: UMIL team