Uncertainty quantification for trustworthy deep learning: Methods and measures

Overview of Uncertainty Quantification

Reliable uncertainty quantification (UQ) is essential for deploying deep neural networks in safety-critical domains like autonomous driving and medical diagnostics. Conventional models often provide overconfident, incorrect predictions, which can lead to catastrophic failures. This paper provides a critical, structured review of UQ methods, organizing them into a taxonomy based on how they generate predictive ensembles and how they measure uncertainty.

Taxonomy of Methods

The authors categorize UQ methods along two primary axes: the generative mechanism (how the predictive ensemble is produced) and the network scope (the portion of the network involved in the uncertainty estimation). The generative mechanisms include:

• Explicit members: Independently trained predictors that are deterministic at inference.
• Posterior sampling: Methods that draw samples from an approximate weight posterior during inference (e.g., Monte Carlo Dropout).
• Closed-form predictive: Analytic methods that provide a predictive distribution in a single forward pass.

By mapping these methods across network scopes—ranging from single deterministic networks to full-network architectures—the authors clarify why certain methods, such as deep ensembles, consistently outperform others in capturing between-mode diversity, while others offer more efficient, albeit approximate, alternatives.

Measuring Uncertainty

A central contribution of this work is the decoupling of the method (which produces the ensemble) from the measure (which summarizes the uncertainty). The authors contrast traditional information-theoretic measures, such as mutual information, with pairwise divergence measures. They emphasize that the choice of measure is as critical as the choice of the underlying model, as different measures capture different aspects of the predictive distribution, such as aleatoric (data-inherent) versus epistemic (model-ignorance) uncertainty.