A three-layer model of natural image statistics

► We propose a three-layer model of natural image statistics. ► After learning, the first layer is selective to Gabor-like features. ► The second layer introduces tolerance to exact localization. ► In the third layer, selectivity to longer contours and texture emerges.