Ria TALUKDER : "Computing with spiking photonic neural networks leveraging sparsity"

Abstract : Spiking neural networks (SNNs) are neuromorphic architectures inspired by biological brains, offering potential advantages in energy efficiency and speed by utilizing sparsity. While CMOS-based electronic implementations of SNNs hold promise, they face challenges in achieving scalability and parallelism. Photonics, on the other hand, provides a powerful alternative, benefiting from the high-speed operation of excitable photonic neurons and the parallel, low-latency nature of optical signal propagation.
In this thesis we introduce a photonic SNN consisting of 40,000 neurons using commercially available optical components such as spatial light modulator and a CMOS camera, providing a scalable and economical solution for photonic SNN proof-of-concept research. A modified Ikeda map governs the dynamics of this system and the introduction of slow and inhibitory feedback gives rise to excitability resembling biological behavior.
Using the liquid state machine (LSM) concept we train our system with Ridge regression and compare its performance to training with the hardware-compatible SPSA algorithm, representing its first application in photonic neural networks. By leveraging latency encoding and sparsity, the network achieves 83.5% accuracy on MNIST with 78.06% sparsity when trained with SPSA and 77.36% accuracy with 62.46% sparsity when trained with Ridge regression.
By integrating photonic nonlinearity, excitability, and sparse computation, this experimental proof-of-concept work lays the foundation for scalable and efficient photonic neuromorphic systems.

Composition du jury :

Daniel BRUNNER, CNRS Researcher, Université Marie et Louis Pasteur, PhD Director

Guy VAN DER SANDE, Professeur, Vrije Universiteit Brussel, Reporter

Damien RONTANI, Assistant professor, Université Paris-Saclay (CentraleSupélec), Reporter

Miguel CORNELLES SORIANO, Tenured Scientist, IFISC, Spanish Research Council, Reviewer

Maxime JACQUOT, Professor, Université Marie et Louis Pasteur, Reviewer

François COURVOISIER, Directeur de recherche, Université Marie et Louis Pasteur, Reviewer

Localization: SupMicrotech ENSMM 26, rue de l'Epitaphe 25000 BESANCON - salle Amphithéâtre Pierre Mesnage