Deep Learning loss landscape: geometry and generalization

Study of the geometrical properties of the loss landscape of growing neural networks: we show that optima are flatter, and improve generalization performances.

Research paper on HAL