I like your likening of MNIST to mouse experiments. Someone should make a hierarchy of evidence equivalent for ML research. Since that's largely focused on medical research.