Neural Network Uncertainty Quantification is the set of methods for estimating the confidence and reliability of neural network predictions — distinguishing between aleatoric uncertainty (irreducible noise in the data) and epistemic uncertainty (model uncertainty from limited training data), enabling AI systems to know what they don't know and communicate confidence levels that are statistically calibrated to actual accuracy rates.

Two Types of Uncertainty

• Aleatoric uncertainty: Inherent noise in the data — cannot be reduced with more data.
• Example: Predicting patient outcome from limited lab values where outcome is genuinely stochastic.
• Modeled by: Predicting output distribution parameters (mean + variance).
• Epistemic uncertainty: Model uncertainty — can be reduced with more training data.
• Example: Model is uncertain about rare drug interactions it rarely saw in training.
• Modeled by: Bayesian posteriors, ensembles, conformal prediction.

Calibration: Expected Calibration Error (ECE)

• Calibration: "When model says 80% confident, is it correct 80% of the time?"
• ECE = Σ (|B_m|/n) × |acc(B_m) - conf(B_m)| where B_m are confidence bins.
• Well-calibrated: ECE ≈ 0. Overconfident: acc << conf. Underconfident: acc >> conf.
• Issue: Modern deep NNs are overconfident — 90% confidence predictions correct only 70% of the time.
• Fix: Temperature scaling (post-hoc): Divide logits by T > 1 → softer distribution → better calibrated.

Monte Carlo Dropout (Gal & Ghahramani, 2016)

• Keep dropout active at inference → stochastic forward passes.
• Run T forward passes with different dropout masks → T predictions.
• Mean of predictions: Point estimate. Variance: Epistemic uncertainty.

model.train()  # keep dropout active
predictions = [model(x) for _ in range(T)]  # T=50 forward passes
mean_pred = torch.stack(predictions).mean(0)
uncertainty = torch.stack(predictions).var(0)
# High variance → high epistemic uncertainty

Deep Ensembles (Lakshminarayanan et al., 2017)

• Train N independent models with different random seeds.
• Predict with all N models → average outputs → variance as uncertainty.
• State-of-the-art for uncertainty estimation; more reliable than MC dropout.
• Cost: N× training and inference overhead.

Bayesian Neural Networks (BNNs)

• Place prior over weights p(W) → compute posterior p(W|data) via Bayes' rule.
• Exact posterior intractable → approximate with variational inference (ELBO).
• Mean-field VI: Factorized Gaussian posterior over all weights → tractable but crude approximation.
• SWAG (Stochastic Weight Averaging Gaussian): Fit Gaussian to trajectory of SGD iterates → practical BNN.

Conformal Prediction

• Distribution-free framework → provable coverage guarantees under mild assumptions.
• Given calibration set: Compute nonconformity scores (e.g., 1 - P(y_true)).
• Set threshold at (1-α)-quantile of calibration scores.
• At inference: Return prediction set C(x) = {y : score(x,y) < threshold}.
• Guarantee: P(y_true ∈ C(x)) ≥ 1-α for any distribution (coverage guaranteed).
• No distributional assumptions → increasingly popular for safety-critical applications.

Out-of-Distribution (OOD) Detection

• Detect inputs far from training distribution → refuse to predict or flag for human review.
• Methods: Maximum softmax probability (simple), Mahalanobis distance, energy score.
• Deep SVDD: Train hypersphere around normal data → distance from center = OOD score.
• Applications: Medical AI refuses prediction on scan from unknown scanner type.

Neural network uncertainty quantification is the epistemic honesty layer that transforms black-box predictors into trustworthy decision support systems — a medical AI that says "I am 95% confident this is benign" when it is only 70% accurate is actively dangerous, while one that correctly identifies its own uncertainty enables clinicians to seek additional tests or expert review exactly when needed, making calibrated uncertainty not merely a technical nicety but the difference between AI that augments human judgment and AI that silently misleads it.