Dynamic linking among neural oscillators leads to flexible pattern recognition with figure-ground separation

Abstract

The basic process of perceptual organization in vision is figure-ground separation. Gestalt psychologists have revealed that this separation is the product of several local coherence criteria. However, relevant information at the global level is also believed to participate in the process, although the neural basis remains unclear. To elucidate the underlying principle of figure-ground separation, we have proposed a neural network model of visual pattern recognition. This model consists of heterogeneous centers composed of nonlinear neural oscillators: a lower center for figure formation, and a higher center for symbol representation. A dynamic linking architecture generates correspondence between the two centers. Pattern recognition is attained as the formation of synchronization of neural oscillators in the two centers. In the present paper, we propose an extended version of our model. Its major improvement is that the synaptic efficacy in dynamic linking is now reversible. Dynamic linking between the two centers occurs via the reversible synapses in the form of transient temporal correlation among neural oscillators. This feature assures further flexibility and diversity of pattern recognition in the present model. Moreover, to elucidate the functional relationships between the local and global levels, bifurcation of the model's behavior was examined in terms of information flow between the two centers. The results suggest that reciprocal information flow between the two heterogeneous centers plays a signiificant role in pattern recognition.