Reproduce “PAC-Bayes bounds for stable algorithms with instance-dependent priors” 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
Omar Rivasplata, Emilio Parrado-Hernandez, John Shawe-Taylor, Shiliang Sun, Csaba Szepesvari
The authors develop a PAC-Bayes bound for a Gaussian-randomized classifier, centering its prior on the learning algorithm’s expected output using stability rather than zero. The theorem relates empirical and population risks, not deterministic SVM test error. Experiments find tighter bounds than an origin-centered alternative in some SVM settings; RIN model selection yielded about 0.12 held-out randomized-classifier error in one 80/20 split.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.