Fig. 2: Artificial neural networks trained to classify naturalistic images show similar patterns of sensitivity as humans.
From: Efficient neural codes naturally emerge through gradient descent learning
a Discrimination thresholds for orientation vary systematically in humans. The sensitivity of the underlying internal representations, as the Fisher Information, can be inferred as the inverse square of the threshold6,9. Data from ref. 44. b We measured the sensitivity of each layer in an artificial network as the change in layer’s response due to a given change in orientation, i.e. the squared norm of the gradient with respect to orientation. c Relative (normalized) sensitivity to orientation for three networks trained on ImageNet, plotted for various layers in each network.
