Epistemic vs aleatoric uncertainty

Summary

Aleatoric uncertainty is irreducible noise in the data-generating process. Epistemic uncertainty is uncertainty about the model itself, due to limited data. More training data shrinks epistemic uncertainty but leaves aleatoric uncertainty unchanged.

Decision-making under uncertainty depends on which kind you have:

High aleatoric, low epistemic: the model knows the world well, but the world is noisy. Collect more diverse features, not more samples. A coin flip has 0.5 aleatoric uncertainty no matter how many flips you observe.
Low aleatoric, high epistemic: the model is uncertain because it has not seen this region of input space. Active learning, more data, ensemble disagreement signals.
Both high: the world is noisy and you have not modeled it well. Both data collection and model improvement help.

Production ML systems that report a single “confidence” number conflate the two and make incorrect downstream decisions: refusing to predict when the world is just noisy, or being overconfident in regions the model has never seen.

How they show up mathematically

For a Bayesian predictive distribution $p (y ∣ x, D) = \int p (y ∣ x, θ) p (θ ∣ D) d θ$ :

Aleatoric: spread within $p (y ∣ x, θ)$ for a fixed $θ$ .
Epistemic: spread of the conditional means $E [y ∣ x, θ]$ over the posterior $p (θ ∣ D)$ .

For regression with a Gaussian likelihood, total variance decomposes as

Var [y ∣ x, D] = aleatoric E_{θ} [σ^{2} (x; θ)] + epistemic Var_{θ} [μ (x; θ)] .

A clean separation: aleatoric is the average noise prediction; epistemic is the disagreement between models.

Learning objective

Does predictive spread live within models or between them?

Read it this way: each line is one model draw's predictive range; each diamond is that model's mean. On the left, individual models are uncertain even though their means agree, so the spread is aleatoric. On the right, each model is confident but their means disagree, so the spread is epistemic. Both panels cover the same overall output envelope, showing why one aggregate “confidence” cannot identify the remedy. This is an original construction based on the law of total variance and Kendall and Gal (2017).

Estimating each in practice

Aleatoric

Predict the noise directly. For regression, output both mean and log-variance: $μ (x), lo g σ^{2} (x)$ . Train with the Gaussian negative log-likelihood (Kendall & Gal, 2017):

L = \frac{1}{2} lo g σ^{2} (x) + \frac{( y - μ ( x ) ) ^{2}}{2 σ ^{2} ( x )} .

The model learns to predict where the data is noisy (large $σ^{2}$ ) and where it is clean (small $σ^{2}$ ). For classification, the predicted probabilities themselves encode aleatoric uncertainty.

Epistemic

Capture model uncertainty. Three practical approaches:

Approach	What it does	Cost
Deep ensembles (Lakshminarayanan et al., 2017)	Train $N$ independent models, look at disagreement	$N$ x training cost
MC dropout (Gal & Ghahramani, 2016)	Keep dropout active at inference, take samples	$N$ x inference cost
Variational Bayes / SWAG / Laplace approx.	Approximate $p (θ ∣ D)$ with a tractable distribution	Training overhead

Deep ensembles are the strongest and simplest; MC dropout is cheaper but a less faithful posterior approximation. Both produce $N$ predictions whose disagreement is the epistemic estimate.

Decomposition

For classification with a deep ensemble of $M$ models producing probabilities $p_{m} (y ∣ x)$ :

H [\overset{p}{ˉ}] = aleatoric \frac{1}{M} m \sum H [p_{m}] + epistemic H [\overset{p}{ˉ}] - \frac{1}{M} m \sum H [p_{m}],

where $\overset{p}{ˉ} = \frac{1}{M} \sum_{m} p_{m}$ is the ensemble average. Predictive entropy splits into within-model (aleatoric) and between-model (epistemic) components.

Where each matters in practice

Active learning: query labels for inputs with highest epistemic uncertainty. Aleatoric uncertainty is irreducible, so labeling a noisy input wastes budget.
Out-of-distribution detection: high epistemic uncertainty signals OOD; high aleatoric does not.
Safe decision-making: medical, autonomous driving. High epistemic uncertainty should trigger “abstain or fall back to human”; high aleatoric uncertainty should still produce a calibrated prediction.
Reinforcement learning: epistemic uncertainty drives exploration (UCB, RND); aleatoric is just reward noise.

Common pitfalls

Reporting “confidence” without saying which kind. A single number conflates the two.
Treating softmax probabilities as a measure of uncertainty. They estimate aleatoric uncertainty but say nothing about epistemic uncertainty. A confidently wrong prediction far from training data has low entropy and high epistemic uncertainty.
Using MC dropout in deployment without retraining the model with dropout active. Dropout-as-Bayes only works if the model was trained with dropout in the same way.
Comparing ensembles of size 2 to “Bayesian deep learning.” Ensembles need 5+ members to capture meaningful posterior spread.