Fig. 3: The sensitivity of ANNs to hue also matches image statistics.
From: Efficient neural codes naturally emerge through gradient descent learning
a The color of a uniform image is varied; in HSV color space, saturation and value are held fixed and the hue is varied. Results are averaged over possible saturations and values. b In humans, the sensitivity to the H axis can be inferred by the perceptual distance between uniformly spaced H values (calculated using the approximately perceptually uniform color space CIELAB) at S=V=1. c The sensitivity to hue in each layer in a trained ResNet18 tracks the empirical frequency of hues in the ImageNet dataset. d Training ResNet18 on a version of ImageNet in which hues are rotated results in a corresponding shift in hue sensitivity.
