In machine learning, knowing what a model doesn't know is often as critical as the prediction itself. For discrete classification tasks, uncertainty quantification (UQ) is relatively straightforward: models output a bounded probability vector over a finite set of classes via a softmax layer, and calibration can be measured by grouping these probabilities into discrete bins.
However, when a model predicts a continuous value—such as a patient's blood pressure, a robot's steering angle, or a molecule's binding affinity—this intuitive framework breaks down. The output space becomes infinite and unbounded. A single scalar prediction offers no inherent signal of its own reliability. As demonstrated by Gustafsson et al., standard deep learning regression models are notoriously poor at self-assessment, frequently exhibiting extreme overconfidence when encountering real-world distribution shifts.
Extending uncertainty quantification to continuous regression requires moving past discrete heuristics and redesigning how models represent, optimize, and calibrate their confidence across continuous manifolds.
The Mathematical Friction of Continuous Spaces
Why can we not simply port classification-based UQ methods directly to regression? The challenge is fundamentally mathematical.
In classification, information-theoretic uncertainty relies on discrete entropy. When moving to continuous variables, we must use differential entropy. Unlike its discrete counterpart, differential entropy can be negative, can grow to infinity, and is not invariant under coordinate transformations. This complicates the derivation of stable, well-behaved uncertainty metrics.
Furthermore, geometric concepts like safety boundaries change completely. As noted by Miri Rekavandi et al., the lack of a discrete decision boundary makes it exceptionally difficult to define a safety radius or certify adversarial robustness for regression models, as there are no clear "label flips" to track. Similarly, Adversarial Bi-Regressor Networks highlight that identifying out-of-distribution (OOD) target samples is significantly harder in continuous spaces because we lack a discrete probability distribution to signal when an input has drifted outside the training domain.
From Prediction Sets to Continuous Intervals
To provide formal statistical guarantees in classification, frameworks like conformal prediction generate a set of likely labels (e.g., {cat, dog}). For continuous spaces, this must be adapted to produce prediction intervals (e.g., $[120.4, 125.8]$) that guarantee the true scalar value falls within the bounds at a user-specified confidence level (such as 95%).
Johansson et al. pioneered this transition by developing specialized nonconformity measures designed explicitly for continuous output spaces. The core objective is efficiency: generating the narrowest possible interval ("sharpness") that still preserves strict statistical coverage.
Classification Conformal Prediction:
Input X ───► [ Model ] ───► Prediction Set: {Cat, Dog} (Discrete labels)
Continuous Conformal Regression:
Input X ───► [ Model ] ───► Prediction Interval: [120.4, 125.8] (Continuous bounds)
To prevent these intervals from degrading under adversarial manipulation or distribution shifts, researchers have integrated structural constraints directly into neural network architectures. For instance, Massena et al. utilize Lipschitz-bounded neural networks to ensure that the predicted intervals remain mathematically stable even under adversarial attacks. To handle covariate shift, Laghuvarapu et al. couple energy-based density estimation with weighted conformal prediction, successfully shrinking the "coverage gap" when predicting continuous properties for out-of-distribution molecules.
Rather than treating conformalization as a post-hoc wrapper, Stutz et al. introduced Conformal Training (ConfTr). By designing a differentiable proxy of the calibration process, the authors allow the network to optimize its parameters end-to-end during training, yielding significantly tighter and more specialized continuous intervals while preserving formal coverage guarantees.
Evidential Learning and Density in Infinite Spaces
Another prominent paradigm is Evidential Deep Learning (EDL), which quantifies epistemic uncertainty (the uncertainty stemming from a lack of training data) in a single forward pass. In classification, EDL places a Dirichlet distribution over the discrete class probabilities.
For continuous targets, placing a meta-distribution over an infinite set of outcomes requires parameterizing continuous probability distributions, such as Normal-Inverse-Gamma distributions. Yoon et al. expanded this capability by introducing flexible distribution models capable of capturing complex, multimodal beliefs over continuous spaces, which is essential when a regression target has multiple plausible solutions. To represent higher-order epistemic uncertainty without relying on expensive ensembles, Wang et al. developed Credal Deep Ensembles, which output probability intervals over continuous densities to explicitly signal model ignorance.
Discrete Evidential Learning:
Model Outputs ───► Dirichlet Distribution Parameters ───► Bounded Class Probabilities
Continuous Evidential Learning:
Model Outputs ───► Normal-Inverse-Gamma Parameters ───► Continuous Probability Density + Epistemic Variance
To ground these evidential models, researchers frequently leverage feature-space density. Charpentier et al. designed Posterior Networks (PostNet), which combine normalizing flows with a Bayesian framework to track data density in the latent space. If a continuous test sample falls into a low-density region, the model automatically scales up its epistemic uncertainty. This approach is supported by the theoretical work of Ji, who formalized how a neural network's empirical error bounds scale directly with training sample density across continuous representation spaces.
Principled Calibration Without Binning
A key challenge in continuous UQ is evaluation: how do we verify that a regression model's error bars actually match reality? In classification, Expected Calibration Error (ECE) relies on partitioning predictions into discrete bins. In regression, binning continuous variances is highly sensitive to hyperparameter choices, introduces statistical bias, and fails to capture the full shape of the error distribution.
To establish a more rigorous foundation, Fishkov et al. extended the theory of proper scoring rules to continuous variables. By deriving closed-form expressions for metrics like the Continuous Ranked Probability Score (CRPS), the authors provide a mathematically sound framework to cleanly decompose predictive regression uncertainty into aleatoric (inherent data noise) and epistemic components without relying on ad-hoc heuristics.
Discrete Calibration (ECE):
| Probabilities | ───► [ Bin 1 ] [ Bin 2 ] [ Bin 3 ] ───► Compare Accuracy vs. Average Confidence
Continuous Calibration (Kernel/Proper Scores):
| Continuous Densities | ───► [ Maximum Mean Discrepancy ] ───► Match Predicted CDF to Empirical Error Distribution
Rather than merely measuring miscalibration after training, Marx et al. reframed calibration as a differentiable distribution-matching problem. By utilizing kernel-based metrics like Maximum Mean Discrepancy (MMD), they enforce that the model's predicted cumulative distribution functions (CDFs) match the true empirical distribution of the residuals. This allows continuous regression models to actively optimize for both sharpness and honest calibration during the training phase itself, bypassing the need for problematic post-hoc binning.