Learning Adaptive Correlations of Independent Components for Complex Cell Modeling

Motivated in part by the hierarchical processing of the cortex, we build an unsupervised network learning the properties of complex cells in V1. Unlike traditional methods, we model the binary relation among these complex cells, which makes our network less constrained and more adaptive for the connectivity among these cells. The obtained filters not only emerge properties similar to those of complex cells, but show more local structures than traditional method such as TICA.