Reproduce “Learning from Aggregate Observations” 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

Yivan Zhang, Nontawat Charoenphakdee, Zhenguo Wu, Masashi Sugiyama

Researchers from The University of Tokyo and RIKEN developed a general probabilistic framework for learning from diverse aggregate observations across multiclass classification and regression tasks. Utilizing maximum likelihood estimation, the framework achieved classification accuracies up to 94.94% on MNIST from triplet comparisons and mean squared error comparable to supervised learning in regression from mean observations.

Key insights

  • The concept of "consistency up to an equivalence relation" is introduced to formally analyze the identifiability and convergence properties of models learned under partial supervision.
  • Classification from pairwise similarity or triplet comparison data is proven to be consistent up to a permutation of classes, indicating that class relationships are learnable even if absolute labels are not unique.
  • Regression from mean observations is consistent (fully recoverable), while regression from rank observations is consistent up to an additive constant, meaning relative values are learnable.

Summary by alphaXiv.

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