Variational Autoencoders (VAEs)

Interactively explore how VAEs encode data into a latent space and reconstruct it, balancing reconstruction and regularization.

VAE Latent Space Simulation

Educational VAE Sandbox

Explanation

  • Encoder: Maps input data to a distribution (mean and variance) in latent space.
  • Decoder: Maps a sampled latent point back to the data space.
  • Reparameterization Trick: Allows backpropagation through random sampling: z = μ + σ * ε
  • Loss Functions: Combines reconstruction loss and KL divergence to regularize latent space.
Drag the orange latent point or click Random Sample z to see how reconstruction changes.
KL divergence is visualized as distance from center.