A neural network implementation of a saliency map model

The saliency map model proposed by Itti and Koch [Itti, L., & Koch, C. (2000). A saliency-based search mechanism for overt and covert shifts of visual attention. Vision Research, 40, 1489–1506] has been a popular model for explaining the guidance of visual attention using only bottom-up information. In this paper we expand Itti and Koch’s model and propose how it could be implemented by neural networks with biologically realistic dynamics. In particular, we show that by incorporating synaptic depression into the model, network activity can be normalized and competition within the feature maps can be regulated in a biologically plausible manner. Furthermore, the dynamical nature of our model permits further analysis of the time course of saliency computation, and also allows the model to calculate saliency for dynamic visual scenes. In addition to explaining the high saliency of pop-out targets in visual search tasks, our model explains attentional grab by sudden-onset stimuli, which was not accounted for by previous models.