Study
Generative AI · Variational autoencoders
A VAE learns a smooth latent space by encoding inputs as distributions and decoding samples from them; a small VAE on handwritten digits, the reparameterisation trick, and generating new digits from noise.
- The encoder outputs a mean and a variance, not a single point, and the KL term keeps the latent space close to a standard normal.
- The reparameterisation trick, z = μ + σ·ε, lets gradients flow through the random sampling step.