The question about delta.

[ghost]

Thanks for your work, it helps me a lot. I notice the delta in your algorithm is gradually accumulate. but that in 《Towards Deep Learning Models Resistant to Adversarial》 is randomly reset. Is there any difference between these two methods ?