Beyond calibration: estimating the grouping loss of modern neural networks

TL;DR

This paper introduces a new estimator for the grouping loss in neural networks, revealing that even well-calibrated models can have significant confidence errors due to grouping loss, especially under distribution shifts.

Contribution

The paper proposes the first estimator for grouping loss in neural networks, emphasizing its importance alongside calibration for reliable confidence scoring.

Findings

Neural networks exhibit notable grouping loss, especially during distribution shifts.

Calibration alone does not ensure accurate confidence scores.

Pre-production validation should consider grouping loss for better reliability.

Abstract

The ability to ensure that a classifier gives reliable confidence scores is essential to ensure informed decision-making. To this end, recent work has focused on miscalibration, i.e., the over or under confidence of model scores. Yet calibration is not enough: even a perfectly calibrated classifier with the best possible accuracy can have confidence scores that are far from the true posterior probabilities. This is due to the grouping loss, created by samples with the same confidence scores but different true posterior probabilities. Proper scoring rule theory shows that given the calibration loss, the missing piece to characterize individual errors is the grouping loss. While there are many estimators of the calibration loss, none exists for the grouping loss in standard settings. Here, we propose an estimator to approximate the grouping loss. We show that modern neural network architectures in vision and NLP exhibit grouping loss, notably in distribution shifts settings, which highlights the importance of pre-production validation.