The Geometric Occam's Razor Implicit in Deep Learning

Benoit Dherin, Michael Munn, David G. T. Barrett

In over-parameterized deep neural networks there can be many possible parameter configurations that fit the training data exactly. However, the properties of these interpolating so…