SGD learning on neural networks: leap complexity and saddle-to-saddle dynamics

Open in new window