TEMPORAL SEGMENTATION OF THE STOCHASTIC OSCILLATOR NEURAL-NETWORK

We propose a stochastic oscillator neural network model of the Hopfiel d-type memory for pattern segmentation tasks exploiting temporal dynam ics of stochastic nonlinear oscillators. The nonlinear oscillators in the model are driven by subthreshold periodic force and noise. For an input pattern which is an overlapped superposition of several stored p atterns, it is shown that the proposed model network is capable of seg menting out each pattern one after another as synchronous firings of a group of neurons. Asystematic study of the dependence on the model pa rameters shows that the temporal segmentation attains its optimal perf ormance at an intermediate noise intensity, which is reminiscent of th e stochastic resonance observed in the coupled oscillator networks. It is also shown that the inhibitory coupling between oscillator groups representing different patterns plays an important role in that it enh ances both the firing rate and the intergroup desynchrony that are ess ential requirements for the optimal performance of the temporal segmen tation.