Denoising MNIST digits

I was wondering how easy (or hard) it would be to build a digit classifier that tells you how to make your digit less noisy as per Jeremy’s description in the lecture. The intuition makes a lot of sense, but my feeling was that taking into account this whole adversarial example business, it might not be that straightforward.

I have trained a simple MNIST classifier (based on the code from examples/ at main · pytorch/examples · GitHub), run some experiments and here’s what I got:

Quite possible I have messed something up, so if anyone want’s to have a go, here’s a: