Event-driven eligibility propagation in large sparse networks

Abstract

Eligibility propagation (e-prop) is a biologically plausible learning rule for recurrent spiking neural networks (SNNs). It approximates gradients computed by backpropagation through time using online updates based on eligibility traces and learning signals. We reformulate the originally time-driven gradient calculations used to compute synaptic weight updates in an event-driven form and show that the resulting formulation accurately reproduces the time-driven results on the benchmark tasks of the original study. We further incorporate additional biological features and show on the event-camera dataset neuromorphic MNIST that they do not degrade learning performance. Enabled by the event-driven formulation, we implement mechanisms that reduce gradient computations, particularly in sparse settings: exactly for the original e-prop model and approximately for the biologically extended model. We assess runtime and memory scaling of both event-driven e-prop models up to millions of neurons, with flat weak-scaling curves suggesting feasibility beyond this range. By providing a reference implementation in the NEST simulation code, an established open-source framework optimized for large-scale SNN simulation, this work supports reproducible studies of e-prop-like learning rules and their prototyping before deployment on neuromorphic hardware.