Spike-based alignment learning solves the weight transport problem.

Abstract

Learning algorithms are often subject to symmetry constraints that are difficult to reconcile with local computation in physical neuronal networks. For example, contrastive Hebbian learning typically assumes symmetric connectivity, while error backpropagation requires knowledge of the forward weights in the backward pass. To solve this weight transport problem, we introduce spike-based alignment learning (SAL), a synapse-local learning rule that harnesses noise for weight alignment. This rule can operate simultaneously with other functional learning rules to maintain the necessary symmetry throughout learning and thereby ensure the correct local representation of gradients. SAL implicitly alleviates any discrepancy arising from the neuron and synapse variability that is ubiquitous in analog substrates, whether biological or artificial. We demonstrate the efficacy of our mechanism using different network models for spiking Bayesian inference and bio-plausible error backpropagation, and benchmark it in a deep learning computer vision task.