Perturbing the convolutional autoencoder

So we have a convolutional autoencoder that works pretty well. Can we make it work even better - either improve the reconstruction accuracy or shrink the encoded state?

• All-convolutional version

The all-convolutional architecture is nice, but the periodic boundary conditions are a nuisance to deal with. We can simplify the autoencoder by pre-scaling the data and rotating the interesting features away from the edges of the field.

• With rotated and scaled training data

To make the autoencoder more useful we can try two ways to shrink the encoded state:

• With an extra strided convolution layer
• With an explicit latent space