Figure 12From: Spatio-chromatic information available from different neural layers via GaussianizationImage manifold transformed through the chromatic layers. Different rows display projections that represent different aspects of the signal, and different columns show how these aspects transform over the series of layers of the network. In the first row the samples are colored according to the actual color they represent (appropriate rendering was done with Colorlab [102]). The scatter plots in the spatio-chromatic rows actually represent 4 dimensional data: the sample in a point in the 3d-axes represents the response of the corresponding chromatic sensor at the three corners of a square, and the color of the sample corresponds to the color of the fourth corner in the square. Correspondence of the color of the sample with the location in the space implies strong correlation between spatial neighbors. Numbers in blue display the 2nd-order correlation measure proposed in [105]: \(C = \frac{1}{2}log_{2}(\frac{\prod_{i} \Sigma _{ii}}{|\Sigma |})\), where Σ is the covariance matrix, and hence C is the difference between the entropy of the marginal distributions and the joint entropy assuming Gaussian approximationsBack to article page