Aleatoric vs epistemic uncertainty revisited

"I reported one number for my doubt. Half of it was the sensor's fault and half was mine, and the pilot could not tell which half to fix."

A Self-Aware AI Agent

Two uncertainties, one predictive distribution

A probabilistic sensor model does not return a number \(\hat{y}\); it returns a distribution \(p(y \mid x)\) over the target given the input window \(x\). Aleatoric uncertainty (from alea, dice) is the spread that remains even with a perfect model trained on infinite data. It is a property of the data-generating process: the same input can map to different outputs because the measurement is noisy or the label is genuinely ambiguous. Why it is irreducible is definitional: no amount of extra training data shrinks the intrinsic scatter of \(y\) around \(x\), because that scatter is baked into how the sensor and the world produce \(y\). Epistemic uncertainty (from epistēmē, knowledge) is the model's uncertainty about its own parameters and structure. It is reducible: it shrinks as data accumulates, and it grows without bound as the input drifts away from anything the model has seen.

The practical test is a thought experiment. Ask: if I collected a thousand more labeled examples exactly like this one, would my uncertainty here go down? If yes, it was epistemic. If the scatter would persist because the target itself is noisy given the input, it was aleatoric. A model near the center of its training distribution can be very epistemically confident and still very aleatorically uncertain, which is exactly the state of a well-trained heart-rate estimator during vigorous motion: it knows the mapping perfectly and the mapping is genuinely blurry.

Aleatoric uncertainty further splits into homoscedastic (constant across inputs, a fixed noise floor) and heteroscedastic (input-dependent). Sensor data is almost always heteroscedastic: a photoplethysmography signal is clean at rest and noisy under motion, a lidar return is sharp on a matte wall and ambiguous on glass. Modeling that input-dependence is the single highest-value move in sensor uncertainty, and we build it explicitly below.

The decomposition: law of total variance

Bayesian modeling makes the split precise. Treat the model parameters \(\theta\) as random with posterior \(p(\theta \mid \mathcal{D})\) learned from data \(\mathcal{D}\). The predictive distribution marginalizes over that posterior:

For a regression target the law of total variance cleaves the predictive variance into exactly the two terms we want:

Read the two terms literally. The aleatoric term is the average, over plausible models, of the noise each model predicts: even after you fix which model is right, this scatter remains. The epistemic term is the variance of the models' mean predictions: if every plausible \(\theta\) agrees on the mean, this term is zero and the model is confident about its own answer; if the posterior contains models that disagree, this term is large. That is why disagreement across an ensemble is a valid epistemic signal, and it is why epistemic uncertainty explodes on out-of-distribution inputs (Chapter 66): far from the training data, nothing pins the parameters down, so the plausible models fan out and their means scatter. For classification the same idea uses entropy: total predictive entropy minus the expected entropy under each \(\theta\) equals the mutual information between the label and the parameters, which is the epistemic (BALD) term.

Modeling aleatoric uncertainty: the variance head

The cheapest genuine win is to let the network predict its own heteroscedastic noise. Instead of a single output \(\hat{\mu}(x)\) trained with mean-squared error, add a second output \(\hat{\sigma}^2(x)\) and train under the Gaussian negative log-likelihood. What this does is let the model say "I am noisier here than there"; why it works is that the NLL loss rewards a large predicted variance exactly where the residual is large, so \(\hat{\sigma}^2(x)\) learns to track the true input-dependent noise; how you implement it is a one-line change to the output layer and the loss, shown next.

import torch, torch.nn as nn

class HeteroscedasticHead(nn.Module):
    """Two outputs per input: predicted mean and predicted log-variance."""
    def __init__(self, in_dim):
        super().__init__()
        self.mu       = nn.Linear(in_dim, 1)
        self.log_var  = nn.Linear(in_dim, 1)   # log-variance: unconstrained, stable

    def forward(self, h):
        return self.mu(h), self.log_var(h)

def gaussian_nll(mu, log_var, y):
    # -log N(y | mu, sigma^2), with sigma^2 = exp(log_var). Constant term dropped.
    inv_var = torch.exp(-log_var)
    return 0.5 * (log_var + inv_var * (y - mu) ** 2).mean()

# One training step on a batch of encoded sensor windows h -> targets y
mu, log_var = head(h)                 # aleatoric: sigma^2 = exp(log_var)
loss = gaussian_nll(mu, log_var, y)   # replaces nn.MSELoss()
loss.backward()

As Listing 18.1 shows, the NLL loss does something MSE cannot: it makes the model attenuate the influence of intrinsically noisy samples instead of straining to fit them, because a large \(\hat{\sigma}^2\) shrinks the \(\exp(-\log\sigma^2)\) weight on that residual. The predicted \(\hat{\sigma}^2(x)\) is a usable, per-window aleatoric estimate you can hand downstream. One caution: this single network still gives you no epistemic term, so it will remain serenely confident on inputs unlike anything it trained on. That gap is precisely what the estimators of Section 18.3 fill.

Why the split changes the decision

The whole point of separating the two terms is that they route to different actions in the system that consumes the prediction (Section 18.7 formalizes how to report them). High aleatoric uncertainty is an instruction to widen margins and keep going; high epistemic uncertainty is an instruction to distrust the model and fall back (Section 18.6). Collapse them and you get both failure directions: you abstain needlessly when the world is merely noisy, or worse, you extrapolate confidently when the model is actually lost.

The when is any sensor system where the cost of a confident wrong answer is high and inputs can leave the training regime: wearables meeting new users, vehicles meeting new weather, industrial monitors meeting a fault mode nobody logged. In those settings the epistemic term is not a nicety; it is the only signal that separates "the world got noisy" from "we left the map."

What's Next

In Section 18.2, we confront a harder truth: even a model that reports the right shape of uncertainty is usually miscalibrated, its stated 90 percent intervals covering the truth far less than 90 percent of the time. We measure calibration on time series, where the usual independence assumptions break, and fix it with temperature scaling and its relatives before we ever reach for the stronger guarantees of conformal prediction.