Reduce load/compute inefficiencies

[bcipolli]

Currently:

• The autoencoder models are loaded twice: once when trying to train them, and again when trying to train the perceptron.
• The hidden unit encodings are always computed, even if the perceptron is cached.

This slows down load time, which can slow down developing and testing new visualizations.