🤖 AI Summary
This work addresses the critical reliability challenge in deep classification known as “high-confidence yet wrong” (JCW) predictions, which undermine abstention and human-in-the-loop mechanisms. It proposes the first training-stage formulation of trustworthy classification as a chance-constrained empirical risk minimization problem, directly bounding the JCW probability by a user-specified risk budget α via a conservative smooth inner approximation. This approach yields a theoretically sound smooth surrogate that enables rigorously certified control over JCW risk. Empirical evaluations demonstrate that the proposed ReliableNet consistently adheres to the prescribed JCW budget across four tabular and two image datasets, achieves the lowest empirical JCW under distribution shift, and excels in accuracy, coverage, calibration, and selective prediction performance.
📝 Abstract
A prediction that is both confident and wrong is a critical reliability failure because it can bypass abstention and human review precisely when the model is mistaken. Empirical risk minimization (ERM) controls average loss but not this failure directly, while calibration, uncertainty estimation, conformal risk control, and selective prediction methods target related reliability properties rather than bounding the joint failure event during training. We propose ReliableNet, which constrains the Joint Confident-Wrong (JCW) probability, the probability that a prediction is simultaneously confident and incorrect, below a user-specified risk budget $α\in(0,1)$. We formulate this as a chance-constrained ERM problem, use a conservative smooth inner approximation whose population feasibility implies the original JCW constraint. Across four tabular and two image datasets, ReliableNet is the only method certified within the JCW budget for every dataset and seed in distribution, when compared against baselines spanning ERM, post-hoc calibration, conformal risk control, and selective prediction. Under demographic, ambiguity, spurious-correlation, novel-class, and covariate shifts, it achieves the lowest empirical JCW among the compared methods while remaining very competitive in accuracy, coverage, calibration, and selective prediction. Risk-coverage results further indicate that ReliableNet achieves better selective ranking than the benchmark methods on most datasets. Overall, ReliableNet provides a principled approach to trustworthy classification.