Quantifying uncertainty in deep learning

Develop reliable and well-calibrated methods to quantify uncertainty in deep learning models so that uncertainty-based active learning strategies can be effective in practice, including for training graph neural networks that predict gene expression responses to perturbations such as GEARS.

Background

In the experimental analysis, the authors evaluate uncertainty-based active learning for training GEARS, a graph neural network that predicts post-perturbation gene expression. They find that an uncertainty-driven selection strategy performs poorly, especially in the early cycles when models are weakly trained and randomly initialized.

They attribute this, in part, to the broader challenge of obtaining trustworthy uncertainty estimates in deep learning, noting that this is a recognized unresolved issue. This motivates the need for principled uncertainty quantification methods that can guide data acquisition decisions reliably in such settings.

References

On the one hand, this stems from the well-known open problem of quantifying uncertainty in deep learning .

— Data Filtering for Genetic Perturbation Prediction  (2503.14571 - Panagopoulos et al., 18 Mar 2025) in Section 4.2 (Results)

Consequently, despite the significant advances brought by deep learning to \ac{doa} estimation, principled uncertainty quantification that remains reliable under challenging operating conditions is still largely an open problem.

— AI-Aided ESPRIT for Joint DoA Estimation and Uncertainty Extraction  (2609.17059 - Zohar et al., 15 Sep 2026) in Section 1, Introduction

These benefits are carried by the posterior's spread, which is exactly the fragile part under amortization, so ``Bayesian'' here is an empirical claim about a calibrated posterior, not a free consequence of emitting a distribution, and validating it --- exact-filter recovery, the amortization gap, and calibration of the posterior precision --- is part of the continuous-Gaussian instantiation we leave to future work (\cref{sec:choice}).

— Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data  (2609.18842 - Hu et al., 16 Sep 2026) in Section 4.3, subsection “The richer belief we do not evaluate” (also discussed in Section 3.2, paragraph “Why a belief and not a point”)

Realizing this loop requires progress beyond predictive accuracy. Key questions include how to calibrate uncertainty under distribution shift, determine when a learned model is reliable enough to guide an experiment, and revise predictive factors as new evidence reveals missing structure.

— JEPA-Anything: Learning Predictive Models across Different Worlds  (2609.20800 - Cui et al., 17 Sep 2026) in Section Future Work

In practice, however, no density estimator or post hoc calibration method yet guarantees a correctly calibrated posterior for real data.

— Strong Lensing Cosmology with Population-level Calibrated Neural Ratio Estimation  (2608.23534 - Jarugula et al., 24 Aug 2026) in Section 1, Introduction, paragraph discussing inference fidelity and trustworthiness

Diffusion models can sample multiple plausible depth maps, but converting these into actionable statistics (e.g., mean and variance) remains an open challenge.

— Monocular Depth Estimation from a Single Image: Progress and Opportunities  (2609.01172 - Liu et al., 1 Sep 2026) in Section 7, “Future Research Directions,” paragraph “High-Fidelity Depth and Uncertainty”