Edge of chaos as a guiding principle for modern neural network training

Lin Zhang, Ling Feng, Kan Chen +1

The success of deep neural networks in real-world problems has prompted many attempts to explain their training dynamics and generalization performance, but more guiding principles…