Encoding the temporal statistics of markovian sequences of stimuli in recurrent neuronal networks

Abstract

Encoding, storing, and recalling a temporal sequence of stimuli in a neuronal network can be achieved by creating associations between pairs of stimuli that are contiguous in time. This idea is illustrated by studying the behavior of a neural network model with binary neurons and binary stochastic synapses. The network extracts in an unsupervised manner the temporal statistics of the sequence of input stimuli. When a stimulus triggers the recalling process, the statistics of the output patterns reflects those of the input. If the sequence of stimuli is generated through a Markov process, then the network dynamics faithfully reproduces all the transition probabilities. © Springer-Verlag Berlin Heidelberg 2002.