Update #3.5

Experiment 1: Consistent t-SNE samples?

Instead of performing t-SNE repeatedly on individual training batches, I decided to perform t-SNE once on the entire training set, and then split them into batches afterwards. In particular, t-SNE was used to find the embeddings on a training set of 20,000 samples, and then the data was split into batches of size [200,500]. Because the embeddings are computed on the whole dataset at once, the mappings are consistent across training batches.

The results are much better!