Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift

TL;DR

This paper evaluates the reliability of predictive uncertainty estimates in machine learning models under dataset shift, revealing that many existing methods fall short but some model-marginalizing approaches perform well.

Contribution

It provides the first large-scale empirical comparison of uncertainty quantification methods under dataset shift in classification tasks.

Findings

Traditional calibration methods are insufficient under dataset shift.

Some model-marginalizing methods maintain strong uncertainty estimates.

Calibration and accuracy degrade with dataset shift across most methods.

Abstract

Modern machine learning methods including deep learning have achieved great success in predictive accuracy for supervised learning tasks, but may still fall short in giving useful estimates of their predictive {\em uncertainty}. Quantifying uncertainty is especially critical in real-world settings, which often involve input distributions that are shifted from the training distribution due to a variety of factors including sample bias and non-stationarity. In such settings, well calibrated uncertainty estimates convey information about when a model's output should (or should not) be trusted. Many probabilistic deep learning methods, including Bayesian-and non-Bayesian methods, have been proposed in the literature for quantifying predictive uncertainty, but to our knowledge there has not previously been a rigorous large-scale empirical comparison of these methods under dataset shift. We present a large-scale benchmark of existing state-of-the-art methods on classification problems and investigate the effect of dataset shift on accuracy and calibration. We find that traditional post-hoc calibration does indeed fall short, as do several other previous methods. However, some methods that marginalize over models give surprisingly strong results across a broad spectrum of tasks.