Recurrent Hamiltonian Echo Learning Enables Biologically Plausible Training of Recurrent Neural Networks

Alice S. Dauphin, Guillaume Pourcel

Keywords: biologically plausible learning, recurrent neural networks, Hamiltonian dynamics, temporal credit assignment, contrastive Hebbian learning, BPTT

Abstract: We combined RHEL, a recently introduced training algorithm, with a Hopfield-inspired Hamiltonian RNN, obtaining a local contrastive Hebbian rule that enables biologically plausible temporal credit assignment and matches BPTT-level performance.

Submission Number: 112