Follow
RT @quocleix@twitter.com
A surprising result: We found that smooth activation functions are better than ReLU for adversarial training and can lead to substantial improvements in adversarial robustness.
http://arxiv.org/abs/2006.14536
