In my last post, I briefly discussed the infuriating fact that a neural network, even when deeply flawed, will often “work” in the sense that it’ll do above-random at classification or a generative network might create things that may sometimes look plausibly from the dataset.
