Consider new features

[thoppe]

Try implementing a convolutional variational autoencoder as a featureset. Use this as a template:

https://github.com/keras-team/keras/blob/master/examples/variational_autoencoder_deconv.py