Implement examples with popular datasets

By some measure, we should be able to provide good performance on classic datasets to show that the pipes are in place to scale to larger examples.

The classic overused dataset everyone uses is [MNIST](http://yann.lecun.com/exdb/mnist/). I'd rather see [fashion-MNIST](https://github.com/zalandoresearch/fashion-mnist) or [EMNIST](https://arxiv.org/pdf/1702.05373v1.pdf).

Good data loading utilities for all of these datasets are with PyTorch via [torchvision.datasets](https://pytorch.org/docs/stable/torchvision/datasets.html).