Separation and Concentration in Deep Networks

John Zarka, Florentin Guth, Stéphane Mallat

Numerical experiments demonstrate that deep neural network classifiers progressively separate class distributions around their mean, achieving linear separability on the training s…