Mechanisms of error-driven learning in cortical neurons

Learning a new skill arises from changes in multiple synapses distributed throughout large networks of neurons. Understanding how this happens is a fundamental question in neuroscience, and critical for developing treatments for disorders of cognition and perception. However, there is currently no consensus on the laws governing synaptic plasticity. Experimental data are typically interpreted in the light of simple Hebbian plasticity, while theoretical models of learning in networks often employ efficient but biologically unrealistic error-backpropagation. We propose that the insights gained from experiments and theory can be reconciled by a radically new view of learning in cortex. We hypothesize that learning is driven by errors arising within neurons themselves; namely that individual dendrites strive to match their neuron’s activity, and that mismatches between somatic and dendritic activity drive plasticity underlying learning. Our preliminary findings show that this principle captures key data on synaptic plasticity, enables effective learning in networks, and can be tested experimentally at the cellular and network level. The proposed programme will combine experimental and computational methods to comprehensively investigate and refine this theory, such that the results will transform our appreciation of the basic building blocks of the brain and provide a formal framework for understanding learning.