A variational autoencoder decodes a latent point z ∈ ℝ² into a data point x ∈ ℝ² with a neural network. Here the decoder (two hidden layers of 32 tanh units, linear output) was trained on the 272 Old Faithful eruptions of Bishop Chapter 9, in standardised units, with a Gaussian likelihood of standard deviation 0.3. All values are in standardised units: each column has mean 0 and standard deviation 1 over the 272 points. Left: the latent plane. Right: the data plane, with the 272 points faint in the background. Click or drag in the left panel to decode a latent point.
The page has no encoder. For the interpolation (preset 3) the page finds, on an 81 by 81 grid over the latent square, the latent point whose decoded output is closest to each of the two eruptions (an approximate inverse, not the trained encoder), then interpolates linearly in latent space and decodes. The manifold readout compares how far the decoded output moves along each latent direction over the prior square. Samples use a seeded generator, so a class sees the same points. The decoder is the trained network stored in code/data/vae_faithful_decoder.json, copied into this page so it works offline.