Modified from the MNIST convolution/deconvolution variational autoencoder example
here. The network demonstrated here is the generative decoder portion (see
Jupyter notebook). The network generates an image through a series of Deconvolution2D layers from coordinates in the 2D latent space. All computation performed entirely in your browser. Toggling GPU on/off shouldn't reveal any significant speed differences, as this is a fairly small network. In the architecture diagram below, intermediate outputs at each layer are also visualized.