Reconstruction loss measures whether a decoder reproduces the input. In a variational autoencoder, it pushes decoded samples toward the original image while regularization shapes the latent distribution. section 17.1.2
A VAE reconstruction pass
An input image becomes a latent sample and then a reconstructed image: z_mean, z_log_variance = encoder(input_img); z = z_mean + exp(z_log_variance) * epsilon; reconstructed_img = decoder(z). Training uses reconstruction loss to make reconstructed_img match input_img. section 17.1.2
Use this rule of thumb:
- Reconstruction loss preserves input information; regularization loss encourages a well-rounded latent distribution and reduces overfitting. section 17.1.2
Figure 17.3 An autoencoder: mapping an input x to a compressed representation and then decoding it back as x'
The autoencoder diagram shows the target relationship: encode an input x, then decode it as a matching x'. section 17.1.2
