VAE ELBO and Reparameterization

From a point to a distribution

Reparameterization

z = mu + std * eps moves randomness outside the learned parameters so gradients can reach mu and std.

Two loss terms

Reconstruction preserves input information; KL divergence keeps the posterior near the prior. Poor weighting can produce weak reconstructions or posterior collapse.

Shape checks

I verified that mu, logvar, eps, and z all use (batch, latent_dim) and made the KL reductions explicit.

Next rerun

The latent-interpolation output was not saved. I will vary β and plot reconstruction against latent continuity in a new run.