Aleatoric and Epistemic Uncertainty in Machine Learning: A Tutorial Introduction

Eyke Hüllermeier, Willem Waegeman

Abstract

The notion of uncertainty is of major importance in machine learning and constitutes a key element of machine learning methodology. In line with the statistical tradition, uncertainty has long been perceived as almost synonymous with standard probability and probabilistic predictions. Yet, due to the steadily increasing relevance of machine learning for practical applications and related issues such as safety requirements, new problems and challenges have recently been identified. In particular, this includes the importance of distinguishing between (at least) two different types of uncertainty, often referred to as aleatoric and epistemic. In this paper, we provide an introduction to the topic of uncertainty in machine learning as well as an overview of attempts so far at handling uncertainty in general and formalizing this distinction in particular.

Introduction: Beyond Single-Score Confidence

Machine learning models are increasingly deployed in safety-critical domains such as medical diagnosis and autonomous driving. A significant risk in these applications is the tendency of models to produce high-confidence predictions even when they are fundamentally incorrect, particularly when encountering out-of-distribution data. Traditionally, models output a single probability score, which conflates two distinct sources of uncertainty: aleatoric and epistemic.

The Core Distinction: Aleatoric vs. Epistemic

Aleatoric uncertainty refers to the irreducible, statistical uncertainty arising from inherent randomness or noise in the data. Examples include predicting a coin flip or dealing with overlapping class distributions where the features provided are insufficient to distinguish between classes.

Epistemic uncertainty, by contrast, is systematic uncertainty caused by a lack of knowledge about the optimal model. This typically occurs when training data is sparse in certain regions of the input space. Crucially, epistemic uncertainty can be reduced by gathering more data, whereas aleatoric uncertainty remains constant regardless of the sample size.

Formalizing Uncertainty in Model Induction

To understand how these uncertainties manifest, we must look at the model induction process. As shown in the following diagram, model induction involves taking background knowledge and training data $\mathcal{D}$ to produce an induced predictor $\hat{h}$ via a learning algorithm.

The process of model induction and prediction, showing the flow from training data to predictions.

The resulting model $\hat{h}$ maps test data $x$ to predictions $\hat{y}$. Uncertainty can enter this pipeline at various stages: through noise in the training data $\mathcal{D}$ (aleatoric) or through the limitations of the induction principle and learning algorithm (epistemic).

Decomposing Uncertainty in the Hypothesis Space

A more formal decomposition involves comparing the induced predictor $\hat{h}$ against the "best possible" hypothesis $h^*$ within a chosen hypothesis space $\mathcal{H}$, and the true "ground truth" function $f^*$.

Decomposition of uncertainty within the hypothesis space, illustrating the gap between the induced model, the best possible model, and the ground truth.

As illustrated above, model uncertainty represents the distance between the ground truth $f^*$ and the best possible hypothesis $h^*$ in our space $\mathcal{H}$. This occurs if our chosen model class is fundamentally incapable of representing the true process. Approximation uncertainty is the distance between $h^*$ and our actual induced predictor $\hat{h}$, often caused by limited data or optimization challenges.

Methodological Landscape

Various frameworks have been developed to quantify these uncertainties:

  • Frequentist Approaches: Methods like maximum likelihood and probability calibration (e.g., Platt scaling) primarily capture aleatoric uncertainty.
  • Bayesian Approaches: By maintaining a distribution over possible hypotheses, Bayesian Neural Networks (BNNs) and Gaussian Processes (GPs) can capture epistemic uncertainty. Practical proxies like Deep Ensembles or Monte Carlo Dropout measure the disagreement between multiple model versions to estimate this lack of knowledge.
  • Set-Based Methods: Standard probability often fails to represent "pure ignorance." Set-based approaches like Conformal Prediction or Credal Sets address this by outputting sets of possible labels or distributions, allowing the model to express uncertainty by increasing the size of the prediction set.

Key Insights and Future Challenges

A critical takeaway is that uncertainty is context-dependent; adding features can turn aleatoric uncertainty into epistemic uncertainty (which can then be resolved). However, evaluating these models remains difficult because datasets lack "ground truth" uncertainty labels. Future work must address "open world" assumptions where models encounter entirely new classes not seen during training.

Summary updated

Abstract

The notion of uncertainty is of major importance in machine learning and constitutes a key element of machine learning methodology. In line with the statistical tradition, uncertainty has long been perceived as almost synonymous with standard probability and probabilistic predictions. Yet, due to the steadily increasing relevance of machine learning for practical applications and related issues such as safety requirements, new problems and challenges have recently been identified by machine learning scholars, and these problems may call for new methodological developments. In particular, this includes the importance of distinguishing between (at least) two different types of uncertainty, often referred to as aleatoric and epistemic. In this paper, we provide an introduction to the topic of uncertainty in machine learning as well as an overview of attempts so far at handling uncertainty in general and formalizing this distinction in particular.
BibTeX
@article{hullermeier2021aleatoric,
  author = {Eyke Hüllermeier and Willem Waegeman},
  title = {{Aleatoric and Epistemic Uncertainty in Machine Learning: A Tutorial Introduction}},
  journal = {Machine Learning},
  year = {2021},
  url = {https://openreview.net/forum?id=u9OK8PsEWnr},
}
Figure