Few-shot pattern detection by transient boosting of somato-dendritic coupling

Neurons are thought to detect salient patterns amidst noise in continuous information streams, but their rapidity tends to be overlooked. Consequently, theoretical neuron models lack key mechanistic features that are suggested to underlie biological neuron rapid learning of input patterns. To unravel these features, we propose a class of models endowed with biologically-plausible predictive learning rules. In these models, an error signal propagates somatic spiking activity to dendrites, facilitating unsupervised learning of repeatedly coactivated presynaptic-neuron communities. Spike-triggered transient boosting of dendritic coupling bestows plausibility and improves the signal-to-noise ratio of learning dramatically. We demonstrate that our plasticity rule enables neurons to swiftly establish a behavioral timescale reward-place association in spatial navigation tasks and showcase how cell assemblies pre-configured in recurrent networks learn multiple patterns within a few repetitions robustly. Our results shed light on the self-supervising function of backpropagating action potentials for pattern learning and its acceleration by pre-existing cell assemblies.