A contractive autoencoder learns representations that change little when similar inputs change slightly, encouraging those inputs to share a latent-space representation. section 9.7
Its role in an autoencoder workflow is to:
- Reduce overfitting through a methodology distinct from the sparsity penalty used by sparse autoencoders. section 9.7
- Support feature learning and dimensionality-reduction applications alongside other autoencoder variants. section 9.11
Conceptual example
Treat two highly similar inputs as deserving nearly the same latent representation rather than encoding their small differences as meaningful features. section 9.7
