Local dendritic balance enables learning of efficient representations in networks of spiking neurons

Significance Neurons have to represent an enormous amount of sensory information. To represent this information efficiently, neurons have to adapt their connections to the sensory inputs. An unresolved problem is how this learning is possible when neurons fire in a correlated way. Yet, these correlations are ubiquitous in neural spiking, either because sensory input shows correlations or because perfect decorrelation of neural spiking through inhibition fails due to physiological transmission delays. We derived from first principles that neurons can, nonetheless, learn efficient representations if inhibition modulates synaptic plasticity in individual dendritic compartments. Our work questions pairwise Hebbian plasticity as a paradigm for representation learning and draws a link between representation learning and a dendritic balance of input currents.