Supervised learning of spike-timing based decision rules

Haim Sompolinsky
Hebrew University, Jerusalem
Israel

Abstract

Neurons in the visual, somatosensory and auditory systems are known to encode stimulus information in the precise timing of action potential. Although computational advantages of temporal codes have long been recognized, it is unknown how neurons can learn to read out spike-timing based representations. Together with Robert Guetig, we have developed a novel biologically plausible supervised synaptic learning rule that enables spiking neurons to learn to decode information embedded in the time domain. The new learning rule - the tempotron, supports storage of several temporal patterns per synapse, roughly matching the capacity achieved in the well-known rate based perceptron learning. Notably, due to the underlying non-linearity in the processing of information across time, our model neuron can learn to discriminate between firing patterns that differ only in their higher-order spike statistics. Our work is the first demonstration that neurons are capable of learning a broad range of spike-time based decision rules, in a manner which is robust against substantial degree of temporal noise.

Return to the Theoretical Neuroscience Symposium Main Page