Title: Estimating Uncertainty via Variance-Gated Distributions

Abstract: Accurately assessing per-sample uncertainty from neural networks is a critical prerequisite for decision-making in high-stakes scenarios. The standard methodology involves leveraging predictive distributions from Bayesian or approximate models, subsequently breaking down the total predictive uncertainty into two distinct parts: epistemic uncertainty, which stems from model limitations, and aleatoric uncertainty, which arises from inherent data noise. However, the validity of this additive decomposition approach has recently faced scrutiny. To address this, we present an intuitive framework for estimating and decomposing uncertainty that relies on the signal-to-noise ratio observed in class probability distributions across various model outputs. Central to our approach is a variance-gated metric that adjusts predictions using a confidence coefficient obtained from ensemble methods. Furthermore, we employ this metric to examine the phenomenon of diversity collapse within committee machines.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC