Stimulus segmentation in a stochastic neural network with exogenous signals

Segmentation by synchrony of firing is investigated in stochastic neural networks with binary neurons. The network is trained by presenting sparsely coded patterns with a Hebbian-type learning rule. Retrieval of these patterns and synchrony of firing is investigated by presenting one or multiple patterns simultaneously to the network. For stimuli consisting of several superimposed patterns the model correctly predicts a high covariance of firing (indicating synchrony) for neurons which are excited by the same pattern in the stimulus. The model gives a negative covariance for neurons which are stimulated by different patterns (indicating absence of synchrony). To obtain useful covariance levels, noise levels should neither be too large, since this would induce complete random firing, nor be too small, since this would minimize the chance to switch between firing assemblies. Furthermore, the strength of the multipattern input may neither be too large, as this would overrule the memory function of the network, nor so small that the desired firing assemblies are never attained.