Machine Learning with Equilibrium Propagation

Abstract

 

In 2017, Scellier and Bengio introduced Equilibrium Propagation, a machine learning algorithm in which the stationary state of a dynamical system performs the computation. A simple illustrative example involves a network of masses coupled by nonlinear springs. Learning is implemented by tuning the spring constants in such a way that the ground state of the network implements the desired input-output map. 

 

I will present the basic principles of the Equilibrium Propagation algorithm and discuss several recent extensions. These include quantum systems, finite temperature systems, the extension to time-dependent Lagrangian dynamics, and the case of nonreciprocal forces. These developments broaden the applicability of Equilibrium Propagation, making new connections between machine learning, quantum mechanics, statistical physics, and dynamical systems theory.

 

 

Biography

 

Serge Massar obtained his PhD from the Université libre de Bruxelles (ULB) in 1995. After postdoctoral positions in Tel Aviv and Utrecht, he returned to ULB in 1998 as a Research Associate of the FRS-FNRS. He is currently Full Professor at ULB.

 

His research interests include quantum gravity, quantum information, quantum and nonlinear optics, and photonic neuromorphic computing. He has co-authored over 200 scientific articles and conference proceedings. His work has been recognized by several awards, including the Gödel Prize in 2023. He is a member of the Royal Academy of Science, Letters and Fine Arts of Belgium.