Continual Learning in a Multi-Layer Network of an Electric Fish

Abstract

Distributing learning across multiple layers has proven extremely powerful in artificial neural networks. However, little is known about how such multi-layer learning is implemented in the brain. Here we provide a detailed account of multi-layer learning in the electrosensory lobe (ELL) of mormyrid fish, a tractable biological model system, and report how two longstanding problems in multi-layer learning are solved. First, we find that the problem of separating learning signals from output signals suitable for driving behavior is solved by a functional compartmentalization, in which inputs driving learning differentially affect dendritic and axonal spikes in intermediate layer neurons. Second, we find that error 'credit assignment' is solved by organizing the connectivity between intermediate and output layer neurons on the basis of learning rather than sensory response selectivity. These mechanisms have broad relevance to learning in the cerebellum, hippocampus and cerebral cortex as well as in artificial systems.