How can a physical system learn through its own dynamics? Equilibrium propagation offers an appealing approach by extracting learning signals from a system’s response to perturbations. Its standard formulation, however, relies on an underlying energy function and does not directly extend to systems with non-reciprocal interactions, like feedforward neural networks.
In this talk, we present Asymmetric EP, which extends equilibrium propagation to non-reciprocal systems. We also introduce a new variational approach to learning through Dyadic EP and Dyadic Backpropagation. This approach connects inference and credit assignment within a common dynamical framework. In particular, they draw on Lagrangian mechanics to recover exact backpropagation gradients.
These results connect learning algorithms with the physics of non-conservative systems and offer new perspectives on how physical systems can learn.
Detalls de contacte:
Miguel C. Soriano Contact form