New arXiv survey maps 5 uncertainty quantification methods for trustworthy deep learning

A new arXiv survey from researchers H. Martin Gillis and Thomas Trappenberg (arXiv:2607.28248) delivers a structured, critical review of uncertainty quantification (UQ) methods for deep learning. As neural networks increasingly power safety-critical systems—from medical diagnosis to autonomous driving—the need for reliable predictive confidence has become urgent. The paper goes beyond traditional deep learning surveys by separating the method that produces a predictive distribution from the measure that summarizes its uncertainty, providing a unified framework for comparison.

The survey organizes UQ methods into five families: Bayesian neural networks, Monte Carlo Dropout, deep ensembles, efficient ensemble approximations, and last-layer or single-pass approaches. For each, the authors examine theoretical motivation, implementation details, empirical performance, and limitations. They also situate adjacent work on evidential networks, prior networks, conformal prediction, and post-hoc calibration, along with decision-time tasks like out-of-distribution detection and selective prediction. The paper then reviews ensemble diversity theory and uncertainty measures, contrasting entropy decompositions with pairwise divergence metrics. It closes with a focused discussion of uncertainty in large language models and open research directions, including efficient epistemic measures for classification, last-layer diversity, and diversity under distribution shift. This is a must-read for ML engineers and researchers looking to build more trustworthy AI systems.