Photonic Neuromorphic Computing

by Serge Massar (Université libre de Bruxelles)

Tuesday 22 Oct 2024, 16:00 → 17:00 Europe/Rome

Aula Conversi (Dip. di Fisica - Edificio G. Marconi)

The biological brain operates in ways that, although poorly understood, are very different from the machine learning and artificial intelligence algorithms implemented today. This suggests that it is possible to process information using physical systems in unconventional ways, that could potentially have benefits in terms of speed or energy consumption. Here we present photonic implementations of two simple machine learning algorithms, namely reservoir computing and extreme learning machines, that are suited to processing time series and for classification, respectively. Concerning reservoir computing, we present photonic implementations based on a simple architecture consisting of a single nonlinear node and a delay line. We also present, for both algorithms, implementations in which the internal variables (the “neurons”) are represented by the amplitudes of a frequency comb. We discuss the advantages and disadvantages of these implementations, how to create an optical output, how to mitigate the effect of noise, how to realize deep architectures in which the output of one system is used as input to the next. We discuss the future of photonic reservoir computing and more generally of analog brain inspired computing.