| Left: A schematic showing how deep neural networks induce simple input / output maps as they become infinitely wide. Right: As the width of a neural network increases , we see that the distribution of outputs over different random instantiations of the network becomes Gaussian. |
