Characterizing Well-Behaved vs. Pathological Deep Neural Networks

Antoine Labatie

We introduce a novel approach, requiring only mild assumptions, for the characterization of deep neural networks at initialization. Our approach applies both to fully-connected and…