A Pulsed Neural Network Model for Attentional Shifts without External Inhibition

摘要

In natural settings such as visual search, the focus of attention should be shifted to find the target. To explain the shifts of visual attention driven by visual stimuli, most models assume the existence of saliency-map, which represents saliency of objects in the visual field. Then, a WTA (winner-takeall) network not only represents the most salient location in the visual environment, but also shifts the focus of attention. The shift mechanism has not been studied well. In this study, we proposed a pulsed neural network model that can move the winning location on the WTA network sequentially. In the traditional model, movement of winner was realized by external Inhibition-of-return(IOR). By contrast, our model used the internal dynamics of saliency map alone, namely combinations of lateral inhibition and self inhibition. A simulation experiment on visual search task showed that performance of the model is consistent with human performance with pop-out and conjunction search tasks. In order to shift the focus of attention, it is not necessary to assume the existence of IOR.