Mr Hugo Tessier from the Mathematical and Electrical Engineering department and Lab-STICC laboratory, will present his research about:
"Convolutional Neural Networks Pruning and its Application to Embedded Vision Systems"
Being at the state of the art in many domains, such as computer vision, convolutional neural networks became a staple for many industrial applications, such as autonomous vehicles—about which Stellantis have ambitions. However, neural networks can bear a great algorithmic complexity, as well as a large memory footprint, which makes them potentially unusable on embedded hardware such as those equipped on such vehicles. In order to reduce this complexity, while keeping the performance that said complexity is supposed to enable, the domain of neural networks compression proposed multiple families of methods, such as pruning that aims at simplifying networks by removing parts deemed unnecessary. Yet, the apparent simplicity of this principle actually hides many subtle implications that have a decisive impact on the efficiency of pruning. In order to clarify the unsuspected complexity of this method and to answer the question of its true efficiency, this manuscript tackles thematically each aspect of pruning and discusses both its theoretical foundations and its practical consequences. It also details the academical and industrial implications of various original contributions of this thesis about parameters supression, layers interdependencies and the energetic efficiency of pruned networks.
Thesis acreditation from IMT Atlantique with the Doctoral School SPIN
Key-words: Computer Vision, Deep Learning, Compression, Pruning