Self-organization of hierarchical visual maps with feedback connections

Visual areas in primates are known to have reciprocal connections. While the feedforward bottom-up processing of visual information has been studied extensively for decades, little is known about the role of the feedback connections. Existing feedback models usually employ hand-coded connections, and do not address how these connections develop. The model described in this paper shows how feedforward and feedback connections between cortical areas V1 and V2 can be learned through self-organization simultaneously. Computational experiments show that both areas can form hierarchical representations of the input with reciprocal connections that link relevant cells in the two areas.