Exploring the Limits of Epistemic Uncertainty Quantification in Low-Shot Settings

TL;DR

This paper evaluates how different uncertainty quantification methods perform in low-shot learning scenarios, revealing that calibration and out-of-distribution detection are highly sensitive to training set size, especially for gradient-based methods.

Contribution

It provides a comprehensive empirical comparison of seven uncertainty methods across varying training set sizes on standard datasets, highlighting their limitations in low-data regimes.

Findings

Calibration error increases with smaller training sets.

Gradient-based methods poorly estimate epistemic uncertainty in low-data settings.

Most methods are miscalibrated on small training sets.

Abstract

Uncertainty quantification in neural network promises to increase safety of AI systems, but it is not clear how performance might vary with the training set size. In this paper we evaluate seven uncertainty methods on Fashion MNIST and CIFAR10, as we sub-sample and produce varied training set sizes. We find that calibration error and out of distribution detection performance strongly depend on the training set size, with most methods being miscalibrated on the test set with small training sets. Gradient-based methods seem to poorly estimate epistemic uncertainty and are the most affected by training set size. We expect our results can guide future research into uncertainty quantification and help practitioners select methods based on their particular available data.