I teach students to build v. small, uncomplicated VAEs on MNIST. No CNNs, just fully-connected layers. We then navigate through the latent space, exploring it

I teach students to build v. small, uncomplicated VAEs on MNIST. No CNNs, just fully-connected layers. We then navigate through the latent space, exploring it by e.g. interpolating between the mean vector for the number 7 and its positive and negative standard deviations. We then decode the latent vectors revealing interesting relations like "the closer you get to one stdev in the positive direction, the more the 7 looks like 9" or "7 is entangled with skew"