GAE
Requires Blocks and Fuel. Modified from vdumoulin's blocks implementation of VAE.
This is a project about making generative autoencoders.
In contrast with VAE, we use deterministic encoder/decoder, and we enforce a prior over the latents by encouraging the encoding of the dataset to appear Gaussian according to some measure.
In practice, we use minibatches instead of the whole dataset.