Louis ANDREOLI : "Spatial recurrent neural networks: from electro-optical towards all-optical"

PhD works of Louis ANDREOLI : "Spatial recurrent neural networks: from electro-optical towards all-optical"

Abstract : Since recent years artificial intelligence and more particularly neural networks play a major role in our technological societies.  Nevertheless, neural networks still remain emulated by traditional computers, resulting in challenging problems such as parallelization, energy efficiency and potentially speed.  A change of paradigm is desirable but implementing neural networks in hardware is a non-trivial challenge.  One promising avenue is optical neural networks, potentially avoiding parallelization bottlenecks.  We experimentally implemented the first system where nodes and connections are implemented in parallel hardware, and we exploit our novel system for unique experiments.  The electro-optical neural network is composed of up to 961 neurons.  Connections between neurons are implemented in fully parallel by optical diffraction, offering highly advantageous scalability.  The network's architecture is based on the principle of reservoir computing, therefore the number of readout weights to be trained is simply the nodes number.  The trainable output layer is realized by a digital micro-mirror device, which, once trained, has the advantage of being fully parallel, passive, and therefore without bandwidth limitation and energy consumption.  Using a textit{Greedy} evolutionary learning, we find that the learning excellently converges.  Furthermore, it appears to possess a conveniently convex-like cost-function and demonstrates exceptional scalability of the learning effort with system size.  Finally, the fundamental bases of an all-optical neural network have been studied.  Neurons will be implemented in an array of quantum dots micropillar lasers optically coupled.  An in-depth study of the optical pumping of micropillar lasers, ultimately allowing a highly scalable system, has revealed a pumping efficiency high enough to realistically operate around a hundred neurons simultaneously.

Jury composition : 

Daniel BRUNNER, CRCN, Université Bourgogne Franche-Comté, PhD Director

Claudio CONTI, Associate Professor, University Sapienza, Reporter

Antonio HURTADO, Professeur des Universités, University of Strathclyde, Reporter

Sylvain GIGAN, Professeur des Universités, Sorbonne Université, Reviewer

Maria-Pilar BERNAL, Directeur de recherche, Université Bourgogne Franche-Comté, Reviewer

Maxime JACQUOT, Professeur des Universités, Université Bourgogne Franche-Comté, PhD co-director

Localization  : 15B, Avenue des Montboucons 25000 BESANCON - salle Amphithéâtre Jean-Jacques Gagnepain