DESNOS Nolan : Nonlinear photonic lanterns for neural networks

Résumé de thèse :

Spiking neural networks (SNN) are bioinspired computing paradigms that enable machine learning with sparse and limited training data. SNNs implemented on photonics based hardware has the potential to realize highly energy−efficient systems (several orders of magnitude). In a novel approach, we investigate the potential of fiber based systems to realize such a neural network based on spikes. Using an architecture analogous to photonic lanterns we aim to jointly address two open challenges; a large scale interconnectivity of the neurons, and signal cascadability through several layers of the network. In the context of this thesis, we envisage the experimental realization of key bricks of this architecture i) A single layer of the network with a sizeable number (~100 ) of physical input and output neurons functionalized with nonlinear optical materials ii) wavelength multiplexing to virtually increase the total number of accessible neurons.

During this thesis, we will work on the characterization of the first generation of the nonlinear photonic lanterns developed at FiberTech Lille, and investigate the fundamental limits to multiplexing and signal cascadability in these systems

Doctorant : DESNOS Nolan

Directeur de thèse : ANDRESEN Esben Ravn