Embedded AI

WP4 focuses on optimisation in artificial intelligence along two main axes: 1) computational limitations and 2) data limitations.

For the first axis, it is a question of focusing on the compression of neural networks of all types in order to reduce both the size of these networks and the inference time they need to produce their predictions.

The second axis is to study architectures that allow learning from multimodal data, 2) unlabelled or weakly labelled data and 3) continuously arriving data, which requires learning throughout the life of the algorithm.

This work package therefore aims at investigating innovative solutions to overcome two major obstacles in current AI: the lack of properly labelled data and the lack of storage and computational capacity on lightweight and embedded systems.

The work package is coordinated by Prof. Thierry Dutoit who works in collaboration with Prof. Sidi Mahmoudi and Dr. Matei Mancas at UMons. A kick-off meeting of WP4 took place on 2 February 2022. During this kick-off meeting, senior researchers as well as newcomers were able to present their work. In addition to launching the work on this research axis now that the milestones of the theses are becoming clearer, this kick-off was also an opportunity for the TRAIL community to get to know each other better. In concrete terms, the researchers of WP4, but also of the other WPs, can better appreciate their respective work through very short presentations in only a few slides. A plan of meetings was then thought to set up a dynamic between the various researchers mobilized around this theme of AI optimization.

Currently, 11 TRAIL affiliated researchers are working on WP4. Among these 11 researchers, 6 PhD students and one post-doc are involved in the ARIAC project.