Uncertainty Estimation

Uncertainty estimation is the practice of having a model report not only a prediction but also how much that prediction should be trusted, in a form that can be acted on: an interval, a probability, a set of plausible answers, or a flag to defer to a human. It matters wherever errors have unequal costs, such as medical diagnosis, autonomous driving, and automated decisions about credit or safety, because a model that is right 95 percent of the time but cannot say which 5 percent is hard to deploy responsibly.

The foundational distinction is between aleatoric and epistemic uncertainty. Aleatoric uncertainty is irreducible noise inherent in the data, such as a blurry image, an ambiguous label, or the outcome of a dice roll, and collecting more data of the same kind does not remove it. Epistemic uncertainty is the model's ignorance, caused by limited training data or an input far from anything seen before, and it can be reduced with more data or a better model. Kendall and Gal (2017) popularized this split in deep learning for computer vision, modelling aleatoric uncertainty with a network that predicts its own noise level and epistemic uncertainty by Bayesian approximations. The split is conceptually useful, though the boundary is blurry in practice, since what counts as noise depends on which features the model is given.

Ensembles are the most reliable practical recipe. Lakshminarayanan, Pritzel and Blundell (2017) showed that training several networks from different random initializations and averaging them gives better calibrated predictions than a single network, with disagreement among members acting as an estimate of epistemic uncertainty. On a confident in-distribution input the members agree; on an unfamiliar input they scatter. The cost is training and serving several models, which is why cheaper approximations are popular. Ovadia et al. (2019) compared several methods under dataset shift and found deep ensembles consistently among the strongest, with uncertainty quality degrading for all methods as the shift grew.

Monte Carlo dropout is the cheap approximation, from Gal and Ghahramani (2016, "Dropout as a Bayesian Approximation"). They showed that a network trained with dropout can be interpreted as approximating a Bayesian model, so keeping dropout active at prediction time and averaging many stochastic forward passes yields a spread that estimates epistemic uncertainty with no change to training. It is easy to adopt, but the quality of the approximation is debated, and in head-to-head comparisons it is often outperformed by ensembles. Other Bayesian methods, such as variational inference and Laplace approximations, aim for the same goal with different trade-offs.

Conformal prediction takes a different philosophy: it makes no claims about the model internals and instead wraps any predictor with a statistical guarantee. Developed by Vladimir Vovk, Alex Gammerman and Glenn Shafer, it uses a held-out calibration set to choose a threshold, so that the prediction set it outputs (a range for regression, a set of labels for classification) contains the true answer with a user-chosen probability such as 90 percent. Angelopoulos and Bates (2021) wrote an accessible tutorial that helped make it popular in machine learning. The guarantee is marginal, meaning it holds on average over the data rather than for every individual input, and it assumes the calibration and test data are exchangeable, so it can fail under distribution shift.

For language models the problem is harder, because a model chooses among many wordings that mean the same thing. Token-level probabilities can mislead when the same answer can be phrased in many ways; Kuhn, Gal and Farquhar (2023) and Farquhar et al. (2024, Nature) proposed semantic entropy, which clusters sampled answers by meaning before measuring their spread, as a way to flag likely hallucinations. A confidence-score read from output probabilities or a verbalized statement is related but not identical: uncertainty estimation is the broader goal, a confidence score is one signal for it, and calibration is the test of whether such numbers mean what they claim.

Common pitfalls follow from over-reading the output. A single softmax probability is not an uncertainty estimate and can be high on nonsense inputs. Ensemble members that share training data and architecture can be confidently wrong together. Uncertainty methods are also hardest to trust in exactly the situations where they are most needed, under distribution shift, so they should be validated on shifted data, not only on the original test set.

Read uncertainty estimation as a model's answer to the question "how sure are you, and why": the aleatoric part says the world is noisy, the epistemic part says the model has not seen enough, and ensembles, MC dropout, and conformal prediction are three different, imperfect ways of turning that humility into a number someone can act on.