Conformalized Bayesian Neural Networks

The primary motivation for this hybrid approach is to fix the overconfident or poorly calibrated uncertainty estimates that arise when a Bayesian model's prior or likelihood is misspecified. While a standard BNN outputs a posterior predictive distribution, this distribution is only as good as the model's assumptions. The conformal wrapper acts as a distribution-free calibration layer that adjusts the size of prediction sets derived from the BNN's posterior to guarantee valid marginal coverage, regardless of how wrong the underlying Bayesian model is. This decouples the quality of the learned representation from the statistical validity of the final output.

The key design choice lies in defining the nonconformity measure using the BNN's posterior. Instead of a single point estimate, the BNN provides a full distribution over predictions. Effective nonconformity scores leverage this richness:

• Posterior Variance: Use the predictive variance as the score, flagging inputs where the model is inherently uncertain.
• Log Predictive Density: Score a point by its negative log-likelihood under the posterior predictive distribution.
• Credible Interval Length: Measure how far a point lies from the center of the highest posterior density interval. This allows the conformal calibration to adaptively widen prediction sets in regions of high epistemic uncertainty.

A core advantage of this hybrid is the ability to distinguish between two fundamental types of uncertainty. The BNN's posterior captures epistemic uncertainty (model ignorance reducible with more data) through weight-space distributions, while the residual variance captures aleatoric uncertainty (inherent data noise). The conformal calibration step then provides a rigorous, finite-sample guarantee on the total prediction set size. This gives engineers a powerful diagnostic tool: they can observe if a wide prediction set is caused by a misspecified prior (epistemic) or inherently noisy data (aleatoric), guiding whether to gather more data or refine the model architecture.

The standard implementation uses the split conformal prediction framework to avoid the prohibitive cost of full Bayesian retraining. The workflow is:

1. Train the BNN on a proper training set to obtain an approximate posterior.
2. For each point in a held-out calibration set, draw multiple posterior predictive samples and compute a nonconformity score (e.g., the negative log predictive density).
3. Compute the empirical quantile of these scores to define a threshold.
4. For a new test point, the prediction set includes all labels whose nonconformity score falls below this calibrated threshold. This process is computationally tractable and provides an exact, distribution-free coverage guarantee.

A practical challenge is the computational cost of Bayesian inference. Conformalized BNNs often rely on approximate inference methods to remain feasible:

• Monte Carlo Dropout: Uses dropout at test time as a variational approximation to the posterior, making sampling extremely cheap.
• Deep Ensembles: Treats a collection of independently trained deterministic networks as a point-mass approximation of the posterior.
• Stochastic Weight Averaging Gaussian (SWAG): Fits a Gaussian distribution to the trajectory of SGD iterates for a low-cost posterior approximation. The conformal wrapper is agnostic to the inference method, providing valid coverage even if the posterior approximation is crude, as long as the calibration and test data are exchangeable.

While conformalized BNNs guarantee marginal coverage (correct coverage on average across all inputs), they do not automatically guarantee conditional coverage (correct coverage for a specific input, e.g., a specific patient or pixel). A misspecified BNN might produce systematically tight prediction sets for one subgroup and wide sets for another, and the conformal wrapper only corrects the average. Achieving approximate conditional validity requires techniques like Mondrian conformal prediction, which applies the calibration independently within pre-defined subgroups, or by using the BNN's posterior to define adaptive, input-dependent nonconformity scores that partially mitigate the issue.

The probability distribution over a new observation after integrating out the model parameters using the posterior. Formally defined as p(y|x, D) = ∫ p(y*|x*, θ) p(θ|D) dθ**, where D is the training data and θ represents the model weights.

In a Conformalized BNN, this distribution serves as the underlying predictive engine that the conformal wrapper calibrates. The raw posterior predictive intervals may be too narrow or too wide due to model misspecification. The conformal step adjusts these intervals using a calibration set to achieve the target coverage rate, effectively correcting the Bayesian credible intervals into frequentist-valid prediction sets.

The condition where the assumed probabilistic model—including the prior distribution and likelihood function—does not contain the true data-generating process. In BNNs, common misspecifications include:

• Prior Misspecification: Choosing a Gaussian prior when the true weight distribution is heavy-tailed
• Likelihood Misspecification: Assuming homoscedastic Gaussian noise when the data exhibits heteroscedasticity or outliers
• Structural Misspecification: Using an insufficiently deep or wide architecture

Misspecification breaks the theoretical guarantees of Bayesian credible intervals. Conformal prediction provides a distribution-free correction layer that restores valid marginal coverage even when the underlying Bayesian model is wrong, making the combined system robust to these failures.

The specific conformalization procedure most commonly applied to BNN outputs. The workflow proceeds in three stages:

1. Training Phase: Fit the BNN on a proper training set to learn the approximate posterior
2. Calibration Phase: Use a held-out calibration set to compute nonconformity scores based on the BNN's posterior predictive distribution. The (1-α) quantile of these scores becomes the threshold
3. Prediction Phase: For each new test point, construct a prediction set by including all labels whose nonconformity score falls below the calibration threshold

This split approach avoids the need to retrain the computationally expensive BNN, making it practical for production deployment while delivering the marginal coverage guarantee: P(Y_test ∈ C(X_test)) ≥ 1-α.

A fundamental decomposition in uncertainty quantification that Conformalized BNNs handle distinctly:

Epistemic Uncertainty (Model Ignorance):

• Arises from limited data or model capacity
• Reducible with more training data
• Captured by the spread of the posterior over weights in a BNN
• Manifests as wide predictive distributions in regions far from training data

Aleatoric Uncertainty (Inherent Noise):

• Arises from irreducible randomness in the data-generating process
• Cannot be reduced with more data
• Captured by the likelihood function's noise parameter

Conformal calibration primarily corrects for epistemic miscalibration, ensuring that the final prediction sets are neither overconfident nor underconfident regardless of how well the BNN separates these two uncertainty sources.

A closely related hybrid approach that applies conformal calibration to deep ensembles rather than BNNs. While BNNs capture uncertainty through weight-space integration, deep ensembles capture it through function-space diversity by training multiple deterministic networks with different random initializations.

Key comparison points:

• BNNs: Theoretically grounded in probability theory, but approximate inference can be biased
• Deep Ensembles: Computationally simpler, often achieve better predictive performance, but lack a formal Bayesian interpretation
• Conformalized Variants: Both benefit from the conformal wrapper, which corrects their raw uncertainty estimates to achieve valid coverage

The choice between them often depends on whether the application requires a coherent Bayesian framework for downstream reasoning or simply needs well-calibrated prediction sets.