I think this sort of thing is handy for reinforcing to people that neural nets are still just code, and should be treated as such. That means testing and reliab

I think this sort of thing is handy for reinforcing to people that neural nets are still just code, and should be treated as such. That means testing and reliability standards built into deployment. The last exercise in particular rams it home really well - you don't need to do adversarial attacks by means of inferring gradients and all that jazz, when you can just spot a buffer overflow. This is a great repo, which I'm (ML Lead) going to make my co-workers read.