Unsupervised Learning

For Distributional Principal Autoencoders we prove a closed-form data score – encoding geometry identity, and that any latent coordinates beyond the data manifold are completely uninformative.