Sparse codes of V1 simple-cells and the emergence of globular receptive fields - a comparative study

Summary Abstract multiple cause models with sparse priors linear or non-linear superposition of basis functions maximization of the data likelihood on image patches likelihood maximization using a novel form of variational EM (ET) same parameter set and training method for both models comparative analysis of the obtained basis functions Results Gabor-like basis functions are obtained in both cases more elongated basis functions when using the non-linear model higher fraction of globular basis functions for the non-linear model