Touching off of this: We have access to the learned encryption function, f : plaintext -> ciphertext.

Most importantly, f is differentiable. It has to be, since it was trained with gradient descent.

So if you want to decipher a ciphertext Y, then use backpropagation to find the X that minimizes distance(f(X), Y). You already know dF/dX (f is differentiable after all), so run this through your magic optimization solver and you'll get an X very close to the result.

Researchers use this "backproagation in input space" idea to recover the training set of a trained neural network, or to debug their model by finding inputs that activate certain neurons, for example. (This is the basic idea behind Deepdream from last year)

It's not even correct to call this "cryptography." The transformation isn't one-way.