Uncertainty in Machine Learning

Hans Weytjens, Wouter VerbekearXiv 2025 · DOI · arXiv

Introduction

Most machine learning models are dangerously overconfident, providing predictions without any sense of how likely they are to be wrong. In high-stakes environments—such as autonomous driving, medical diagnosis, or credit scoring—a wrong prediction can be catastrophic. The fundamental issue is that standard machine learning (ML) models are designed to always produce an output, even when the input is completely unlike anything seen during training. They lack a "reject option": the ability to abstain from a prediction when confidence is low.

This research provides a rigorous framework for distinguishing between uncertainty caused by a lack of data (epistemic) versus inherent noise (aleatoric), offering practical methods to implement a "reject option" in business-critical systems. For AI to be truly useful in business, it must move beyond point estimates and provide "uncertainty-aware" predictions. This allows organizations to build dual-track systems where high-confidence cases are automated, while low-confidence cases are routed to human experts.

Methodology for Quantifying Uncertainty

The paper explores uncertainty quantification across three major model families and introduces a robust, model-agnostic framework:

  • Linear Regression: Uses classical Ordinary Least Squares (OLS) to derive analytical confidence intervals. While straightforward, it relies on strict assumptions like homoscedasticity (constant noise across the domain).
  • Random Forests: Leverages information theory. It uses Shannon entropy to measure total uncertainty, conditional entropy for aleatoric uncertainty (noise), and mutual information to isolate epistemic uncertainty (model knowledge gaps).
  • Neural Networks: Focuses on Bayesian Neural Networks (BNNs). Since true BNNs are computationally intractable, the authors use Monte Carlo Dropout as an approximation. They also implement loss attenuation, where the network learns to predict its own variance ($\sigma^2$) to handle heteroscedastic noise (noise that varies across the data).
  • Conformal Prediction: This is considered the "gold standard" for practical applications. It is a model-agnostic wrapper that uses a small calibration dataset to generate statistically valid prediction sets (for classification) or intervals (for regression) with a user-defined confidence level (e.g., 90%).

Key Insights: Epistemic vs. Aleatoric Uncertainty

A vital distinction made in the research is between epistemic and aleatoric uncertainty.

  1. Epistemic Uncertainty: This is reducible by adding more data; it represents gaps in the model's knowledge.
  2. Aleatoric Uncertainty: This is irreducible noise, inherent to the system (like the outcome of a coin flip).

Knowing which type of uncertainty is dominant dictates the organizational strategy: either collect more data to fill knowledge gaps or invest in better sensors to reduce noise. Furthermore, the authors highlight the "Fallacy of Linearity." Linear models often fail to signal when they are extrapolating. OLS confidence intervals might not widen enough when moving far outside the training domain, leading to a false sense of security.

Experimental Results

The authors demonstrate these methods using several benchmarks, including the Iris dataset, the Covertype dataset, and simulated sinusoidal data.

BNN Robustness and Epistemic Visualization

In the sinusoidal regression task, BNNs successfully "averaged out" Gaussian noise in high-variance regions, providing a good fit while correctly flagging those regions as high-uncertainty. The visualization below shows how epistemic uncertainty spikes in regions where training data is absent.

Visualization of epistemic uncertainty on a simulated sinusoidal dataset, highlighting regions where the model lacks training data.

Conformal Validity and Business Impact

On the Covertype dataset, the authors set a target confidence level of 80%. The resulting conformal prediction sets contained the true target in 80.35% of test cases, demonstrating the framework's mathematical rigor.

From a business perspective, the paper demonstrates that by setting a total uncertainty threshold (e.g., 0.9), a business can guarantee a specific accuracy level (e.g., 97%) by only automating the subset of cases (e.g., 86%) that fall below that threshold. This is evaluated using Accuracy-Rejection Curves, where model performance improves as the most uncertain predictions are removed.

Limitations and Future Work

While promising, these methods face several challenges:

  • Computational Cost: Monte Carlo Dropout requires multiple forward passes during inference, which may be too slow for real-time applications.
  • The Confidence-Utility Trade-off: In Conformal Prediction, higher confidence levels inevitably lead to larger prediction sets. If a 99% confidence set includes every possible class, it is statistically valid but practically useless.
  • Data Drift: While rising uncertainty can signal data or concept drift, these methods are primarily ex-post. Integrating these into the training loop to optimize for both accuracy and uncertainty simultaneously remains a complex area of ongoing research.

Summary updated

Abstract

This book chapter introduces the principles and practical applications of uncertainty quantification in machine learning. It explains how to identify and distinguish between different types of uncertainty and presents methods for quantifying uncertainty in predictive models, including linear regression, random forests, and neural networks. The chapter also covers conformal prediction as a framework for generating predictions with predefined confidence intervals. Finally, it explores how uncertainty estimation can be leveraged to improve business decision-making, enhance model reliability, and support risk-aware strategies.
BibTeX
@misc{weytjens2025uncertainty,
  author = {Hans Weytjens and Wouter Verbeke},
  title = {{Uncertainty in Machine Learning}},
  year = {2025},
  eprint = {2510.06007},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.06007},
}
Figure