Surrogate Gap Guided Minimiation Improves Sharpness-Aware Training

ICLR 2022

Juntang Zhuang [1], Boqing Gong [2], Liangzhe Yuan [2], Yin Cui [2], Hartwig Adam [2], Nicha C. Dvornek [1], Sekhar Tatikonda [1], James S. Duncan [1], Ting Liu [2]

[1] Yale University

[2] Google Research

Lower curvature leads to better generalization

Comparison of different training schemes

Validations on toy examples and deep-learning experiments

Citation

@inproceedings{

zhuang2022surrogate,

title={Surrogate Gap Minimization Improves Sharpness-Aware Training},

author={Juntang Zhuang and Boqing Gong and Liangzhe Yuan and Yin Cui and Hartwig Adam and Nicha C Dvornek and sekhar tatikonda and James s Duncan and Ting Liu},

booktitle={International Conference on Learning Representations},

year={2022},

url={https://openreview.net/forum?id=edONMAnhLu-}

}