.@reinerpope and I talked about how neural nets are kind of like cryptographic ciphers in reverse, and how one of the biggest attacks against ciphers is related to gradient descent.