Training neural networks is one of the defining computational problems of our time. This motivates the search for physical systems that can directly identify optimal model weights through the natural dynamical evolution of the system. Oscillator networks are an exciting platform for such physical computers, having demonstrated unique speed and energy advantages in tasks such as combinatorial optimization and time series classification. However, current progress towards training neural networks using oscillator-based systems have invariably relied on external control electronics that update system parameters following a learning rule. In my talk, I will present our recent efforts towards fully physical training in oscillator networks. To accomplish this, we co-locate weights and activations in a single physical system, using stable attractors to realize long-term memory for the weights, and short-lived degrees of freedom for the activations. Weight updates are implemented through the nonlinear dynamical evolution of the network, which realizes a local learning rule. I will discuss recent numerical results highlighting how relatively simple inter-modal interactions can give rise to emergent self-learning, and comment on our ongoing efforts to experimentally demonstrate physical training in a platform of MEMS oscillators.

Marc Serra-Garcia is a tenure-track group leader at the AMOLF institute on the physics of complex matter in Amsterdam, the Netherlands — A position he took after his studies at Caltech and ETH Zurich. His research focuses on developing elastic structures with novel functionalities by combining advanced design algorithms and precision microfabrication with insights from fundamental physics. His work includes the invention of a nonlinear system that extracts energy from random vibrations, the demonstration of wave-controlling materials based on topological insulators, and, currently, the development of elastic computing structures — from speech classifiers to entirely-mechanical microprocessors; area of research that has recently been recognised by an ERC Starting grant.