Incremental Self-Organization of Spatio-Temporal Spike Pattern Detection

Abstract

Nervous systems utilize temporally precise patterns of activity. However, the mechanisms by which spike patterns are processed are not known. In particular, the fact that during learning different patterns are distributed over time raises the question of how groups of neurons become selective for new spike patterns without overwriting already learned patterns. A simple one-layer spiking neural network model is presented that learns to recognize spatiotemporal spike patterns sequentially. The approach integrates biological synaptic mechanisms, including Hebbian learning, heterosynaptic plasticity, and synaptic scaling, allowing groups of neurons to self-organize selectivity for a set of spike patterns. Spoken words, transformed by a cochlear model into spatio-temporal spike patterns, are learned without supervision. This work suggests how the brain can use temporal spike codes and provides a novel, scalable, efficient, and noise-tolerant solution to the stability-plasticity dilemma.