Beyond gradients: Factorized, geometric control of interference and generalization

Abstract

Interference and generalization, which refer to counter-productive and useful interactions between learning episodes, respectively, are poorly understood in biological neural networks. Whereas much previous work has addressed these topics in terms of specialized brain systems, here we investigated how learning rules should impact them. We found that plasticity between groups of neurons can be decomposed into biologically meaningful factors, with factor geometry controlling interference and generalization. We introduce a "coordinated eligibility theory" in which plasticity is determined according to products of these factors, and is subject to surprise-based metaplasticity. This model computes directional derivatives of loss functions, which need not align with task gradients, allowing it to protect networks against catastrophic interference and facilitate generalization. Because the models factor structure is closely related to other plasticity rules, and is independent of how feedback is transmitted, it introduces a widely-applicable framework for interpreting supervised, reinforcement-based, and unsupervised plasticity in nervous systems.