The Data Manifold under the Microscope

Marios Koulakis, Constantin Seibold

A significant gap exists between theory and practice in deep learning. Generalization and approximation error bounds are often derived for simplified models or are too loose to be…