Towards Better Generalization via Distributional Input Projection Network

Yifan Hao, Yanxin Lu, Hanning Zhang +2

As overparameterized models become increasingly prevalent, training loss alone offers limited insight into generalization performance. While smoothness has been linked to improved…