Phase Collapse in Neural Networks

Florentin Guth, John Zarka, Stéphane Mallat

Deep convolutional classifiers linearly separate image classes and improve accuracy as depth increases. They progressively reduce the spatial dimension whereas the number of channe…