Two-Layered Physiology-Oriented Neuronal Network Models that Combine Dynamic Feature Linking via Synchronization with a Classical Associative Memory

Abstract

A visual object ‘pops out’ against the background, if its local features have a high degree of perceptual coherence: the object’s local features are linked by the visual system into a perceptual whole. A neuronal mechanism for such preattentive ‘object definition’ is likely to be based on stimulus-induced synchronized activities, as observed in the visual system of the cat [l, 2]. Based on neurophysiological evidence we developed a model neuron with the following properties [7]: 1. Convergent feeding connections can directly activate the model neurons spike discharge. 2. Modulatory action of linking inputs onto feeding inputs provides mutual enhancement and synchronization of stimulus-specific responses. The linking effect is represented by the synchronization of the spike trains of model neurons coupled by linking connections. Spatial and temporal continuity of a stimulus, thus, can produce synchronization in those neural assemblies that are activated by such stimulus. The response of a model of two identical layers is similar to that of cortical assemblies in our neurophysiological observations: Stimulated input regions that share some stimulus features induce synchronized ensemble activities via the laterally projecting recurrent linking connections, even if the stimuli are moving and are composed of spatially separated subregions. In a second type of simulation the associative property of the linking network is supported by a further layer with a classical type of associative memory [10–12]. Presenting a stimulus pattern (which is similar to stored patterns), the degree of association to one of these patterns depends on the spatial overlap of the patterns and the strength of the local linking connections in the lower layer.