Reproduce “Domain Generalization via Invariant Feature Representation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Krikamol Muandet, David Balduzzi, Bernhard Schölkopf

DICA learns a low-dimensional representation across patient domains, bringing input distributions closer while retaining information about input-output relationships. A classifier can use it on an unseen patient without adapting to that patient. In leave-one-out evaluation over 30 subjects, accuracy was 94.80 ± 3.81% versus 92.74 ± 5.01% for UDICA with distributional SVM, conditional on stable or smoothly varying input-output relationships.

Key insights

  • Probability distributions can be effectively represented and compared using kernel mean embeddings and a novel distributional variance metric in Reproducing Kernel Hilbert Spaces.
  • The central subspace, which captures the functional relationship between input and output variables, can be preserved by maximizing a term related to the covariance operator inverse regression in RKHS.
  • A derived learning-theoretic bound demonstrates that reducing both inter-domain distributional variance and the complexity of the learned transformation directly leads to improved generalization performance on unseen domains.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.