CUPID: A Plug-in Framework for Joint Aleatoric and Epistemic Uncertainty Estimation with a Single Model
Xinran Xu, Xiuyi Fan
TL;DR
CUPID is a versatile plug-in module that estimates both aleatoric and epistemic uncertainty in pretrained models without retraining, enhancing transparency and trustworthiness in AI applications.
Contribution
It introduces a general-purpose, layer-wise uncertainty estimation method that is model-agnostic and does not require modifications or retraining of the base model.
Findings
Consistently competitive uncertainty estimation across tasks
Provides layer-wise insights into uncertainty sources
Applicable to classification, regression, and out-of-distribution detection
Abstract
Accurate estimation of uncertainty in deep learning is critical for deploying models in high-stakes domains such as medical diagnosis and autonomous decision-making, where overconfident predictions can lead to harmful outcomes. In practice, understanding the reason behind a model's uncertainty and the type of uncertainty it represents can support risk-aware decisions, enhance user trust, and guide additional data collection. However, many existing methods only address a single type of uncertainty or require modifications and retraining of the base model, making them difficult to adopt in real-world systems. We introduce CUPID (Comprehensive Uncertainty Plug-in estImation moDel), a general-purpose module that jointly estimates aleatoric and epistemic uncertainty without modifying or retraining the base model. CUPID can be flexibly inserted into any layer of a pretrained network. It…
Peer Reviews
Decision·ICLR 2026 Poster
+ CUPID generally outperforms or matches baselines like MC Dropout, Rate-in, PostNet, BNN, DEC, and BayesCap, with its two branches complementing each other across tasks. + A theoretical section provides a first-order Taylor expansion showing that epistemic uncertainty scales with the product of network sensitivity and feature deviation.
- The AU branch resembles standard heteroscedastic regression, and EU relies on output-preserving perturbations, which is close in spirit to prior work that estimates distributional shift via reconstruction error [RUE (Wang et al., 2023)]. It is unclear what the most significant novelty of this paper is. - CUPID adds a learned module plus an extra forward of the perturbed path, but there is no computational cost analysis. - The authors claim CUPID produces reliable uncertainty estimates, howeve
**1. Well-written:** The paper is well-written and well-motivated with good literature review. **2. Experimental results:** The paper is enriched with numerous experimental evaluations of their methods in different tasks and demonstrated their success with AUC, AURC, spearman and Pearson coefficients. They also evaluated the calibration error. Their proposed method perform better in almost all cases.
**1.Confusing interpretation of epistemic uncertainty:** The epistemic uncertainty has been considered as the discrepancy between the original and the perturbed prediction. However, why this discrepancy captures the model's lack of knowledge is not clear from the paper. **2. Limited novelty in aleatoric uncertainty:** The estimation of aleatoric uncertainty seems similar with BayesCap [2] and therefore that part has limited novelty. [1] Depeweg, S., Hernandez-Lobato, J.M., Doshi-Velez, F.
The proposed CUPID framework offers a simple yet flexible plug-in design that can be applied to pretrained models without retraining. This practical modularity makes the method attractive for real-world applications where uncertainty estimation must be added post hoc. The joint decomposition of aleatoric and epistemic uncertainties provides a unified view of model and data uncertainty, and the layer-wise formulation enables interpretable analysis of which network components contribute most to un
**Methods** The proposed methods for estimating aleatoric and epistemic uncertainty are conceptually straightforward and largely aligned with prior works. The aleatoric branch essentially performs feature-based variance regression, similar to heteroscedastic uncertainty modeling in Kendall & Gal (2017) and Evidential Deep Learning (Sensoy et al., 2018). The epistemic uncertainty is derived from feature reconstruction errors, an idea previously explored in autoencoder-based OOD detection (An & C
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsAdversarial Robustness in Machine Learning · Explainable Artificial Intelligence (XAI) · Machine Learning in Healthcare