Reproduce “Some theoretical improvements on the tightness of PAC-Bayes risk certificates for neural networks” 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
Diego García-Pérez, Emilio Parrado-Hernández, John Shawe-Taylor
The authors train a distribution over models to minimize a PAC-Bayes upper bound on posterior-average 0–1 risk, using a prior chosen independently of training data. They optimize the certificate directly and estimate non-differentiable 0–1 risk from a differentiable surrogate. On CIFAR-10, a 4-layer CNN with 9,946 parameters achieves 0.738 ± 0.007, while deeper models remain vacuous without data-dependent priors.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.