CoLT: The conditional localization test for assessing the accuracy of neural posterior estimates

2025-07-22

AI Summary

This work addresses the problem of rigorously validating the approximation accuracy of a neural posterior estimator (q( heta mid x)) relative to the true posterior (p( heta mid x)). We propose Conditional Localization Testing (CoLT), a hypothesis testing framework that identifies—via a learnable localization function—the regions of the input space where posterior discrepancy is most pronounced, enabling efficient, density-free, single-sample testing. CoLT operates within a simulation-based inference paradigm, avoiding reliance on classifiers or explicit density modeling, thereby ensuring both statistical rigor and computational tractability. Experiments demonstrate that CoLT significantly outperforms existing baselines in detection sensitivity, precision of misfit localization, and interpretability. By providing a principled, scalable, and interpretable diagnostic tool, CoLT establishes a new paradigm for trustworthy validation of Bayesian deep learning models.

Technology Category

• Machine Learning: Calibration & Uncertainty Quantification
• Reasoning under Uncertainty: Relational Probabilistic Models
• Intelligent Robots: State Estimation

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Economics, Online Markets and Human Computation: Data quality aspects of human-annotated datasets

Abstract

We consider the problem of validating whether a neural posterior estimate ( q(θmid x) ) is an accurate approximation to the true, unknown true posterior ( p(θmid x) ). Existing methods for evaluating the quality of an NPE estimate are largely derived from classifier-based tests or divergence measures, but these suffer from several practical drawbacks. As an alternative, we introduce the emph{Conditional Localization Test} (CoLT), a principled method designed to detect discrepancies between ( p(θmid x) ) and ( q(θmid x) ) across the full range of conditioning inputs. Rather than relying on exhaustive comparisons or density estimation at every ( x ), CoLT learns a localization function that adaptively selects points $θ_l(x)$ where the neural posterior $q$ deviates most strongly from the true posterior $p$ for that $x$. This approach is particularly advantageous in typical simulation-based inference settings, where only a single draw ( θsim p(θmid x) ) from the true posterior is observed for each conditioning input, but where the neural posterior ( q(θmid x) ) can be sampled an arbitrary number of times. Our theoretical results establish necessary and sufficient conditions for assessing distributional equality across all ( x ), offering both rigorous guarantees and practical scalability. Empirically, we demonstrate that CoLT not only performs better than existing methods at comparing $p$ and $q$, but also pinpoints regions of significant divergence, providing actionable insights for model refinement. These properties position CoLT as a state-of-the-art solution for validating neural posterior estimates.

Problem

• Validating accuracy of neural posterior estimates
• Detecting discrepancies between true and estimated posteriors
• Adaptively identifying regions of significant divergence

Innovation

• Introduces Conditional Localization Test (CoLT)
• Learns adaptive localization function for discrepancies
• Offers rigorous guarantees and practical scalability