Reproduce “Addressing Failure Prediction by Learning Model Confidence” 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

Charles Corbière, Nicolas Thome, Avner Bar-Hen, Matthieu Cord, Patrick Pérez

ConfidNet learns a score to predict the probability assigned to the true class, rather than using the classifier’s maximum class probability, to rank likely errors. It trains on labeled examples and predicts that score without test labels. On CIFAR-10 with VGG-16, it raises AUPR-Error from MCP’s 45.36% to 49.94%, showing improved error ranking in this setting, not improved classification accuracy.

Summary by alphaXiv.

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