Study and Design of an Energy Efficient Perception Module Combining Event-based Image Sensors and Spiking Neural Networks using 3D Integration Technologies

Abstract
Event-based image sensors can be useful is several scenarios, especially where classical image sensors tend to fail, e.g., in the dark while undergoing fast vibrations, or through a fast movement of a head. They present High Acquisition Speed (HAS), High Dynamic Range (HDR), and Event-Based (EB) acquisition properties. However, they generate an exceptional amount of data, with a substantial quantity of noise in it. In his Ph.D. Study, Maxence tried to understand how EB imagers can be used for computer vision applications, and what are their limits. After having understood these limits, he proposed to implement a Convolutional Spiking Neural Network (CSNN) hardware accelerator for near sensor operation with 3D IC Technologies, following the path laid out by the RETINE sensor and its macropixel readout strategy. This neuromorphic accelerator was optimized for ultra-low power operation and demonstrates state of the art energy efficiency.


Jury
Mr Gilles Sicard, HDR Engineer, CEA LETI, Thesis Director
Mr Michel Paindavoine, Professor, Univ. Bourgogne Franche-Comté, Reviewer
Mr David Bol, Associate Professor, K.U. Leuven, Reviewer
Mrs Loréna Anghel, Professor, Grenoble Institute of Technology and Management, Examiner
Mr Alexandre Valentian, Ph.D. Engineer, CEA LIST, Thesis Supervisor, Invited
Mrs Edith Beigné, HDR Engineer, Facebook, Invited
Mr Christophe Posch, Ph.D. Engineer and CTO, Prophesee, Invited