General Uncertainty Estimation with Delta Variances

arXiv stat.ML / 5/1/2026

💬 OpinionIdeas & Deep AnalysisModels & Research

共有:

Key Points

The paper proposes Delta Variances, a set of algorithms designed to quantify epistemic uncertainty efficiently when decision-makers have limited data.
Delta Variances can be applied not only to neural networks but also to more general functions built from neural-network components.
The authors demonstrate the method on a weather simulator that uses a neural-network-based step function, achieving competitive empirical performance while requiring only a single gradient computation.
The work presents several theoretical derivations of Delta Variances, showing that known uncertainty estimation methods appear as special cases, and it introduces an extension that improves empirical results.
A key practical advantage is that the approach does not require changes to the neural network architecture or the training procedure, making it easy to implement.

Abstract

Decision makers may suffer from uncertainty induced by limited data. This may be mitigated by accounting for epistemic uncertainty, which is however challenging to estimate efficiently for large neural networks. To this extent we investigate Delta Variances, a family of algorithms for epistemic uncertainty quantification, that is computationally efficient and convenient to implement. It can be applied to neural networks and more general functions composed of neural networks. As an example we consider a weather simulator with a neural-network-based step function inside -- here Delta Variances empirically obtain competitive results at the cost of a single gradient computation. The approach is convenient as it requires no changes to the neural network architecture or training procedure. We discuss multiple ways to derive Delta Variances theoretically noting that special cases recover popular techniques and present a unified perspective on multiple related methods. Finally we observe that this general perspective gives rise to a natural extension and empirically show its benefit.