Reproduce “Variational Bayesian Last Layers” 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
James Harrison, John Willes, Jasper Snoek
The authors introduce VBLL, which places a probability distribution over output-layer weights while keeping network features deterministic. It trains this layer using minibatch-estimated lower bounds rather than sampling through the training loss. On CIFAR-10, D-VBLL reports 96.4% accuracy and 0.969 SVHN OOD AUROC, outperforming SNGP on these measures but not on ECE, NLL, or CIFAR-100 near-OOD AUROC.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.