Playing with variational autoencoders

Traversing the latent space on MNIST. On the left, the TSNE 2d embedding of the much larger VAE latent space (6272 elelment vector). I have create a trajectory in the latent space that traverse all digits. The trajectory pass through the median vector of each blob. The red point indicates the position in the embedding of the current latent vector beign decoded.