Figure 4From: Spatio-chromatic information available from different neural layers via GaussianizationLeft: Redundancy reduced at different layers of the network (labels in the horizontal axis) considering samples over the whole image manifold (global experiment). Redundancy is measured in total correlation (in bits/sensor). Lines in different colors refer to different models, and different line styles (solid/dashed) refer to the empirical RBIG estimate given by Eq. (5) and the theoretical estimate given by Eq. (9). Each result accumulates the reductions due to previous layers. Right: Redundancy reduced at the V1 layer (last layer) using different variations of the baseline model: more flexible and more rigid divisive normalization and totally rigid (totally linear) model. In both plots (left and right), error bars stand for the standard deviation over the 10 realizations of the estimation. The theoretical estimation also has error bars because the univariate entropies in Eq. (9) were empirically estimatedBack to article page