Tracing sources of epistemic uncertainty in deep learning predictions: homo- and hetero-scedastic linearized estimators

2026-08-07

AI Summary

This work addresses the challenge of disentangling epistemic uncertainty from uncertainty induced by local data sparsity in deep learning, which are often conflated and difficult to distinguish. The authors integrate two classical statistical uncertainty estimation techniques into a deep learning framework, constructing homoscedastic and heteroscedastic linearized estimators based on an approximate Fisher information matrix. This approach enables, for the first time, a scalable and fine-grained decomposition of epistemic uncertainty and the effects of local data scarcity in modern deep neural networks. Experimental results demonstrate that the proposed method effectively quantifies the relative influence of each uncertainty type across individual test samples, substantially enhancing model robustness and reliability in real-world applications.

Abstract

We adapt two classical statistical estimators for quantifying uncertainty to modern deep learning, in order to provide clearer insights into uncertainty attributable to two sources : aleatoric uncertainty, or locally scarce data. Our approach leverages recent advances in approximate Fisher Information Matrices, to enable scaling to actual architectures. Experimental results demonstrate how each test points is differentially impacted by both sources, highlighting the practical utility of our estimators in improving the robustness of real-world applications.

Problem

• epistemic uncertainty
• aleatoric uncertainty
• deep learning
• uncertainty quantification
• Fisher Information Matrix

Innovation

• epistemic uncertainty
• aleatoric uncertainty
• Fisher Information Matrix
• heteroscedasticity
• uncertainty quantification