Uncertainty Quantification: Predictive Variance, MC Dropout, and Deep Ensembles

The UQ literature for deep nets splits along a methodological axis (how uncertainty is produced) and an operational axis (what guarantee is offered).

Bayesian methods (BNNs, deep ensembles, MC dropout, Laplace) produce a posterior predictive whose moments decompose into epistemic and aleatoric components (Kendall and Gal 2017; Lakshminarayanan et al. 2017). The cost is extra compute and the usual approximation concerns covered in Part III. Prior networks (Malinin and Gales 2018) reach the same decomposition from the other side, learning a distribution over categorical distributions directly and buying a single-forward-pass score in exchange for an auxiliary OOD training set.

Post-hoc calibration methods (Platt scaling, isotonic regression, temperature scaling (Guo et al. 2017)) are nearly free, a single scalar applied to the softmax logits. They correct the marginal confidence histogram and leave the ranking untouched. How much correction a modern network needs turns out to depend on its architecture family and pretraining scale (Minderer et al. 2021), and every fitted calibrator degrades under distribution shift (Ovadia et al. 2019).

Conformal prediction (Vovk et al. 2005; Angelopoulos and Bates 2021) sidesteps the probabilistic modeling question entirely and instead delivers a distribution-free finite-sample coverage guarantee at a target level $1 - \alpha$¹. The trade-off is that individual set sizes can be large if the base model is poor.

OOD detection is the sister literature that reuses these signals as scores for a binary "in-distribution vs. out-of-distribution" classifier (Hendrycks and Gimpel 2017; Liang et al. 2018; Liu et al. 2020), with a feature-space branch that scores distance from class prototypes rather than logit shape (Lee et al. 2018).

Probabilistic programming languages (Pyro, NumPyro, TFP, Stan) (Bingham et al. 2019; Phan et al. 2019; Tran et al. 2018; Carpenter et al. 2017) are the infrastructure layer that ties posteriors, calibrations, and predictive distributions into one model. Bayesian RL (Ghavamzadeh et al. 2015; Osband et al. 2016; Depeweg et al. 2018) applies the decomposition to a dynamics model and uses the epistemic component as an intrinsic reward driving Thompson sampling and information-directed exploration.

Table tab:uq-sota summarizes the main method families and the metrics each one reports.

Three methodological observations follow from the table. First, post-hoc calibration dominates accuracy-matched cost: one extra scalar parameter (the temperature $T$) reduces ECE by roughly an order of magnitude on CIFAR-100 at zero accuracy cost² (Guo et al. 2017), making temperature scaling the default first step for any deployed classifier. Second, OOD detection and calibration are separate problems: a perfectly calibrated model can still fail catastrophically on inputs from outside the training distribution (ImageNet-trained ResNet assigns confident softmax to texture noise), which is why the OOD literature designs a different score (energy, ODIN, likelihood-ratio) rather than reading off softmax confidence. Third, conformal prediction is orthogonal to both: it wraps any base model and converts its score into a set with exact marginal coverage, regardless of whether the score was calibrated or how the model was trained.

Historical Notes

The separation of reducible from irreducible uncertainty predates deep learning by centuries. Laplace's Essai philosophique sur les probabilités (1814) distinguishes between "ignorance of the causes" (epistemic) and "inherent variability" (aleatoric) in the context of classical mechanics. Knight's Risk, Uncertainty, and Profit (1921) made the same distinction in economics, calling the epistemic component true uncertainty and the aleatoric component risk. In machine learning, the distinction was re-introduced by (Neal 1996) in the context of Bayesian neural networks: the posterior variance over weights is epistemic, the likelihood variance (observation noise) is aleatoric. Modern deep-learning UQ is thus not a new paradigm but a practical engineering of ideas that have been known for generations.

Calibration as a desideratum goes back at least to Brier's (1950) meteorological scoring rule, which rewarded reliable probabilistic forecasts of rain. Platt (1999) (Platt 1999) was the first to apply calibration to machine-learning classifiers (SVMs). The "deep learning is overconfident" observation by (Guo et al. 2017) was partly a rediscovery of the same phenomenon that had been noted in boosted trees and had been partly understood as a general property of margin-maximizing classifiers.

Conformal prediction originated in (Vovk et al. 2005), who developed the theory for online sequential prediction with exchangeability. The framework was largely unknown outside the learning-theory community until (Angelopoulos and Bates 2021) wrote an accessible tutorial for deep-learning practitioners, triggering rapid adoption. The tutorial paper became the most-cited introduction to conformal prediction within three years of publication, showing that exposition can be as impactful as novel theory in moving a field.

OOD detection has its roots in the anomaly detection literature (Chandola et al. 2009 survey) but became a deep-learning-specific research program with (Hendrycks and Gimpel 2017), who framed it as a binary classification task with a specific evaluation protocol (CIFAR-10 vs. SVHN, TinyImageNet vs. LSUN). The AUROC/FPR-at-95-TPR metrics became the de facto OOD benchmark³.

References