Interesting, but given how good image classifiers have become I think people are better off inverting the train/test specifications on the regular MNIST set to

Interesting, but given how good image classifiers have become I think people are better off inverting the train/test specifications on the regular MNIST set to push the limits of what we can do with smaller training sets.