7 Learned folds

R. Heller · Date: 2026-07-29 · Last changed: 2026-11-09

Can a network learn the fold and the unfold at once?

Chapter status: planned. The items below describe work still to be implemented; this chapter does not provide a trained autoencoder or benchmark results.

  • A torch autoencoder with a two-unit bottleneck. (keras would pull TensorFlow and break clean-clone reproducibility.)
  • The only method here that learns an explicit inverse: encoder = low-dimensional representation, decoder = reconstruction in the input space.
  • Reconstruction loss is the fold–unfold identity made computable.
  • Compare the learned decoder against the true unfolding, not just against the input.
This book was built by the bookdown R package.