TL;DR
Machine learning models often fail when encountering data unlike their training set (out-of-distribution detection). Existing scoring methods rely on logits or distances with unrealistic assumptions.
✦ Why It Matters
Engineers can deploy more reliable ML systems by using ConjNorm for detecting anomalous inputs before they cause failures.
Key Takeaways
Full Summary
Deployed machine learning systems must detect when input data differs significantly from training data to avoid unreliable predictions. Prior approaches estimate detection scores from model logits (final layer outputs) or geometric distances, but these methods either misrepresent true data density or impose impractical constraints on data distribution assumptions.
ConjNorm proposes a unified theoretical framework grounded in Bregman divergence—a mathematical measure of difference between probability distributions—that extends to exponential family distributions (a broad class of statistical models). The method reformulates density estimation as finding an optimal norm coefficient through a conjugation constraint.
To overcome computational barriers in calculating partition functions (normalization constants), the authors developed an unbiased estimator using Monte Carlo importance sampling, a statistical technique for approximating integrals. Experiments on CIFAR-100 and ImageNet-1K benchmarks demonstrate ConjNorm outperforms existing methods by up to 13.25% and 28.19% on false positive rate metrics, establishing new state-of-the-art performance.
Related