Neural coding and decoding with spike times

The question of how neurons encode, store, and process information has challenged systems neuroscience for more than five decades. Approaches to neuronal coding have long been dominated by the rate coding hypothesis, which stipulates that neurons encode information by the number of action potentials generated in response to a stimulus. Indeed, neuronal firing rates have been used as the central measure for characterizing neuronal responses and much of our intuition of how neurons integrate information arriving from afferent populations is based on these measures. However, the observed fine temporal modulation of the neuronal firing in several systems has motivated the study of alternative schemes of neuronal codes, such as synfire chains (Abeles, 1991; Diesmann et al., 1999) that exploit spike synchrony and rank order decoders (Thorpe and Gautrais, 1998) that are based on spike latencies. Nevertheless, the computational capacity of these schemes remained unknown, and more generally it was unclear to what extent neurons are capable of decoding information that is encoded in the relative timing of incoming spikes. To address this question, the Tempotron has been devised as a generic model of spike-timing-based information processing by single neurons. The Tempotron modifies its synaptic efficacies to carry out peform appropriate classification of incoming spatio temporal patterns of spikes. Using this model, we have shown that a simple biologically plausible form of post synaptic integration suffices to decode a broad range of spike-timing-based neuronal codes with high capacity. In addition, the model clarifies the nature of synaptic learning rules that would allow neurons to realize such capabilities.