[Paper Review] Pitfalls of In-Domain Uncertainty Estimation and Ensembling in Deep Learning

Arsenii Ashukha, Alexander Lyzhov+2 more|arXiv (Cornell University)|Feb 15, 2020

Adversarial Robustness in Machine LearningComputer Science53 references111 citations

TL;DR

The paper analyzes in-domain uncertainty estimation for image classification, highlights pitfalls of common metrics, and introduces the deep ensemble equivalent (DEE) score to compare ensembling methods, showing deep ensembles often outperform others and that test-time data augmentation can substantially boost performance.

ABSTRACT

Uncertainty estimation and ensembling methods go hand-in-hand. Uncertainty estimation is one of the main benchmarks for assessment of ensembling performance. At the same time, deep learning ensembles have provided state-of-the-art results in uncertainty estimation. In this work, we focus on in-domain uncertainty for image classification. We explore the standards for its quantification and point out pitfalls of existing metrics. Avoiding these pitfalls, we perform a broad study of different ensembling techniques. To provide more insight in this study, we introduce the deep ensemble equivalent score (DEE) and show that many sophisticated ensembling techniques are equivalent to an ensemble of only few independently trained networks in terms of test performance.

Motivation & Objective

• Assess reliability and comparability of in-domain uncertainty metrics for DNNs.
• Evaluate a broad set of ensembling techniques on standard image classification benchmarks.
• Introduce calibrated log-likelihood and DEE score to compare ensembles fairly across datasets.
• Investigate the impact of test-time data augmentation on uncertainty estimation and calibration.

Proposed method

• Survey common in-domain uncertainty metrics (LL, Brier score, calibration metrics) and reveal their pitfalls.
• Propose calibrated log-likelihood as a fair metric by evaluating at optimal temperature.
• Define and compute the deep ensemble equivalent (DEE) score to compare ensembling methods to deep ensembles.
• Evaluate a suite of ensembling techniques (deep ensembles, SSE, FGE, SWAG, VI, K-FAC Laplace, dropout, TTA) across CIFAR-10/100 and ImageNet.
• Analyze test-time data augmentation (TTA) as a baseline and its interaction with calibration and temperature scaling.
• Quantify performance using calibrated log-likelihood and DEE across architectures (VGG16, PreResNet, WideResNet, ResNet50).

Experimental results

Research questions

• RQ1 What are the limitations of standard in-domain uncertainty metrics when comparing different models?
• RQ2 How do various ensembling techniques compare in terms of predictive performance and calibration on in-domain data?
• RQ3 Can a unified, interpretable metric (DEE) fairly rank ensembling methods across datasets and architectures?
• RQ4 What is the role of test-time data augmentation in improving in-domain uncertainty estimation and calibration?

Key findings

• Many common in-domain uncertainty metrics (log-likelihood, Brier score, calibration metrics) are not reliably comparable across models and can mis-rank methods.
• Calibrated log-likelihood (evaluated at the optimal temperature) provides a more stable, comparable measure of in-domain uncertainty.
• The deep ensemble equivalent (DEE) score allows comparing ensembling methods by the equivalent size of a deep ensemble; deep ensembles often outperform other methods for a given test-time budget.
• Test-time data augmentation (TTA) substantially boosts ensemble performance, especially on ImageNet, and can rival larger ensembles with minimal cost.
• Methods that explore multiple loss-landscape modes (deep ensembles, snapshot ensembles, cyclical SGLD) outperform single-mode methods such as dropout, VI, and K-FAC Laplace in DEE terms on CIFAR and ImageNet; TTA further enhances calibration and accuracy with proper temperature scaling.
• SSE and cSGLD explore loss-landscape modes effectively, but their gains depend on hyperparameters and dataset; DEE helps diagnose when saturation or underfitting occurs.