You'll notice that the results of our VAE model are a fair bit fuzzier than our standard autoencoder:
In some cases, it also appears to be undecided between several different digits, like in the following example, where it appears to be getting close to decoding to a 7 instead of a 9:
This is because we have specifically enforced the distributions to be close to each other. If we were to try to visualize this on a two-dimensional plot, it would look a little bit like the following:
You can see from this last example that it can generate several different variations of each of the handwritten digits, and also that there are certain areas in between the different digits where it appears to morph between several different digits.
