Sparse Deep Predictive Coding captures contour integration behavior of the early visual system

Both neurophysiological and psychophysical experiments have pointed out the crucial role of feedback and recurrent connections to integrate contextual and attentional modulations in V1 neural activity. While numerous models have accounted for feedback effects at either neural or representational level, none of them were able to bind those two levels of analysis. Is it possible to describe feedback effects at both neural and representational levels using the same model? We answer this question using the Sparse Deep Predictive Coding (SDPC) model that combines Predictive Coding (PC) and Sparse Coding (SC) into a hierarchical and convolutional framework. PC describes the interactions between layers using feedforward and feedback connections, and SC models the internal recurrent processing of each layer. We trained a 2-layered SDPC on two different databases, and we interpret it as a model of the early visual system (V1 V2). We first demonstrate that once the training is achieved the SDPC exhibits oriented and localized receptive fields. Second, we analyze the feedback effects on neural organization in our V1-model using interaction maps. The observed interaction maps reflect the Gestalt principle of good continuation through association fields. We demonstrate that feedback signal aligns neurons in the side-zone of the interaction map co-linearly to the central preferred orientation. In addition, feedback signal excites neurons in the end-zone, and inhibits those located in the side-zone. It suggests that feedback signals promote contour integration in V1. Third, we demonstrate at the representational level that the SDPC feedback connections are able to denoise blurred images. We finally discuss our results in the light of evidences from neuroscience.