01. Understanding Predictive Uncertainty

• Primary References:

1. 1
   Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On Calibration of Modern Neural Networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), PMLR 70:1321–1330.
2. 2
   Kendall, A., & Gal, Y. (2017). What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? In Advances in Neural Information Processing Systems (NeurIPS 2017).
3. 3
   Depeweg, S., Hernandez-Lobato, J. M., Doshi-Velez, F., & Udluft, S. (2018). Decomposition of Uncertainty in Bayesian Deep Learning for Efficient and Risk-Sensitive Learning. In ICML 2018.
4. 4
   Niculescu-Mizil, A., & Caruana, R. (2005). Predicting Good Probabilities with Supervised Learning. In ICML 2005.

• Why I Wrote This Note: When we trained convolutional classifiers for tuberculosis screening from chest radiographs and early hospital readmission prediction for diabetes patients, I noticed that standard cross-entropy training pushes logits to extreme magnitudes. A raw softmax score of \(0.96\) does not mean the patient has a \(96\%\) empirical probability of readmission. I compiled this note to establish a clean mathematical and PyTorch foundation that separates confidence, dispersion (entropy), knowledge uncertainty (epistemic), and calibration.

• Confidence is Pointwise; Entropy is Distributional: Maximum softmax probability \(\hat{p} = \max_c p(y=c\mid x)\) only looks at the winning class, whereas predictive Shannon entropy \(\mathcal{H}[p(y\mid x)]\) integrates dispersion across all \(C\) classes.
• Total Uncertainty Decomposes Cleanly: Under a posterior distribution over parameters \(p(\theta \mid \mathcal{D})\), total predictive entropy decomposes additively into Aleatoric Uncertainty (expected data entropy) and Epistemic Uncertainty (mutual information \(\mathcal{I}[y; \theta \mid x, \mathcal{D}]\)).
• Modern Deep Nets Are Overconfident by Construction: Negative log-likelihood (NLL) continues to decrease even after classification error hits \(0\%\) on the training set, driving weight norms and logit magnitudes upward and causing severe miscalibration.
• Calibration \(\neq\) OOD Detection: Post-hoc Temperature Scaling optimizes a single scalar \(T > 0\) on a validation split to dramatically lower \(\text{ECE}\) in-distribution, while leaving the relative class ranking \(\arg\max_c z_c\) strictly invariant.

4.1 Confidence vs. Predictive Entropy

Let \(f_\theta: \mathcal{X} \to \mathbb{R}^C\) denote a neural network mapping an input \(x \in \mathcal{X}\) to a logit vector \(\mathbf{z} = f_\theta(x) = (z_1, \dots, z_C)^\top\). The categorical predictive distribution is obtained via the softmax transformation:

From this probability simplex \(\Delta^{C-1}\), we extract two distinct scalar summaries:

1. 1
   Maximum Softmax Confidence (\(\hat{p}\)):

1. 1
   Predictive Shannon Entropy (\(\mathcal{H}\)):

To compare entropy across tasks with different class cardinalities \(C\), I normalize entropy by the uniform distribution bound \(\log C\), yielding \(\tilde{\mathcal{H}}(x) = \mathcal{H}[p_\theta(y\mid x)] / \log C \in [0, 1]\). Consider a \(10\)-class problem where Model A predicts \((0.50, 0.50, 0, \dots, 0)\) and Model B predicts \((0.50, 0.055, \dots, 0.055)\). Both share identical confidence \(\text{Conf}(x) = 0.50\), yet Model A exhibits two-class ambiguity (\(\mathcal{H} \approx 0.693\text{ nats}\)) while Model B spreads its remaining mass across nine classes (\(\mathcal{H} \approx 1.792\text{ nats}\)). Entropy captures the full geometry of the tail.

4.2 Information-Theoretic Decomposition of Uncertainty

A single deterministic network cannot distinguish between an intrinsically ambiguous input (e.g., an occluded chest X-ray boundary) and an unfamiliar input from an underrepresented subpopulation. When we maintain a distribution over parameters \(q(\theta) \approx p(\theta \mid \mathcal{D})\) (via an ensemble of \(M\) models or \(M\) stochastic forward passes), the posterior predictive distribution is:

Taking the Shannon entropy of \(\bar{p}(y \mid x)\) and subtracting the expected entropy of individual models yields the Mutual Information (\(\text{BALD}\) objective):

flowchart LR
    X["Input Sample x"] --> ENS["Posterior Samples {θ₁, ..., θ_M}"]
    ENS --> PBAR["Posterior Predictive p̄(y|x)"]
    PBAR --> TU["Total Entropy H[p̄(y|x)]"]
    TU --> AU["Aleatoric: E_θ[H[p(y|x,θ)]]<br/>Irreducible Data Noise"]
    TU --> EU["Epistemic: I[y; θ | x]<br/>Model Ignorance / OOD"]

4.3 Probabilistic Calibration and ECE

A classifier is perfectly calibrated if its predicted confidence matches its true empirical accuracy at every confidence level \(p \in [0, 1]\):

Because \(\hat{P}\) is a continuous random variable, we approximate this expectation in practice by partitioning predictions on \(N\) test samples into \(B\) equally spaced confidence bins \(\{I_1, \dots, I_B\}\) where \(I_b = (\frac{b-1}{B}, \frac{b}{B}]\). Letting \(B_b\) denote the indices of samples whose confidence falls into bin \(I_b\), the Expected Calibration Error (\(\text{ECE}\)) is:

where empirical bin accuracy and mean bin confidence are defined as:

For safety-critical clinical settings, I also track Maximum Calibration Error (\(\text{MCE}\)):

4.4 Post-Hoc Calibration via Temperature Scaling

The simplest and most effective parametric post-hoc calibration method is Temperature Scaling (Guo et al., 2017). Given frozen validation logits \(\mathbf{z}_i\), we rescale by a single positive scalar temperature \(T > 0\):

We optimize \(T^\star\) by minimizing the Negative Log-Likelihood on a held-out calibration split \(\mathcal{D}_{\text{val}}\) via L-BFGS:

As \(T \to \infty\), \(\hat{\mathbf{q}}_i\) approaches the uniform distribution \((1/C, \dots, 1/C)\) with maximum entropy \(\log C\). When \(T > 1\) (typical for overconfident ResNets and Vision Transformers), Temperature Scaling softens the probability distribution and increases predictive entropy without altering top-\(1\) accuracy.

import torch
import torch.nn as nn
import torch.nn.functional as F

class TemperatureScaler(nn.Module):
    # Post-hoc temperature scaling module optimized via L-BFGS on validation logits
    def __init__(self, init_temp: float = 1.5):
        super().__init__()
        self.log_temperature = nn.Parameter(torch.tensor([init_temp]).log())

    @property
    def temperature(self) -> torch.Tensor:
        return self.log_temperature.exp()

    def forward(self, logits: torch.Tensor) -> torch.Tensor:
        return logits / self.temperature

    def fit(self, val_logits: torch.Tensor, val_labels: torch.Tensor) -> float:
        optimizer = torch.optim.LBFGS([self.log_temperature], lr=0.05, max_iter=60)
        criterion = nn.CrossEntropyLoss()

        def closure():
            optimizer.zero_grad()
            loss = criterion(self.forward(val_logits), val_labels)
            loss.backward()
            return loss

        optimizer.step(closure)
        return float(self.temperature.item())

def compute_ece_and_uncertainty(probs: torch.Tensor, labels: torch.Tensor, n_bins: int = 15):
    # Compute Expected Calibration Error (ECE) and normalized Shannon entropy
    confidences, predictions = probs.max(dim=-1)
    accuracies = predictions.eq(labels).float()

    # Normalized predictive Shannon entropy in [0, 1]
    num_classes = probs.size(-1)
    entropy = -(probs * probs.clamp_min(1e-12).log()).sum(dim=-1) / torch.log(torch.tensor(float(num_classes)))

    bin_boundaries = torch.linspace(0.0, 1.0, n_bins + 1, device=probs.device)
    ece = torch.zeros(1, device=probs.device)
    for b in range(n_bins):
        in_bin = (confidences > bin_boundaries[b]) & (confidences <= bin_boundaries[b + 1])
        prop_in_bin = in_bin.float().mean()
        if prop_in_bin.item() > 0:
            acc_in_bin = accuracies[in_bin].mean()
            conf_in_bin = confidences[in_bin].mean()
            ece += torch.abs(acc_in_bin - conf_in_bin) * prop_in_bin
    return float(ece.item()), entropy

In my published work on diabetes patient early readmission risk minimization (IEEE BECITHCON 2024) and tuberculosis screening from chest X-rays (IEEE ICEEICT 2024), raw classifier probabilities cannot be handed directly to clinicians as risk percentages.

• In clinical risk scoring, an uncalibrated model with \(\text{ECE} > 0.12\) systematically distorts downstream expected utility thresholds.
• Furthermore, in TIG-UNet++ lung region segmentation (ICICT 2025), pixel-wise predictive entropy maps highlight boundary ambiguity along the costophrenic angles and diaphragm, giving us a spatial uncertainty mask that explains where the segmentation backbone needs thresholded guidance.
• Finally, calibrated predictive entropy serves as the halting signal in my ongoing experiments on Early-Exit Neural Networks and Risk-Constrained Selective Prediction.

• Label Smoothing vs. Temperature Scaling: Training with Label Smoothing (\(\epsilon = 0.1\)) implicitly prevents logit explosion and reduces pre-scaling \(\text{ECE}\), but it often erases subtle epistemic distance signals in logit space. I have found that training with standard cross-entropy + L2 weight decay followed by post-hoc Temperature Scaling preserves better out-of-distribution separability.
• Adaptive Binning (ACE): Standard fixed-width \(\text{ECE}\) leaves mid-confidence bins (\([0.4, 0.7]\)) nearly empty for high-accuracy classifiers, introducing high variance. Using equal-mass (quantile) bins via Adaptive Calibration Error (\(\text{ACE}\)) provides a much more stable diagnostic on small medical validation cohorts (\(N < 1,000\)).

• Implement reusable TemperatureScaler and compute_ece_and_uncertainty utilities for PyTorch classification and segmentation pipelines.Verified
• Verify pixel-level entropy heatmaps on chest X-ray lung masks in the TIG-UNet++ evaluation pipeline.Verified
• Compare fixed-width \(\text{ECE}\) against Adaptive Calibration Error (\(\text{ACE}\)) and Classwise-ECE on imbalanced clinical tabular datasets.Pending
• Benchmark input-dependent parameterization \(T(x) = \text{Softplus}(w^\top h(x))\) (Parameterized Temperature Scaling) under synthetic corruption shifts.Pending

"A neural network that achieves 99% accuracy on a test benchmark can still be dangerously uncalibrated if its 99% confidence predictions are only correct 80% of the time."

Core Calibration & Decomposition Identity Card:

• Connected Research Notes:
• Deep Ensembles vs MC Dropout — How to instantiate \(q(\theta)\) in practice to compute the epistemic mutual information \(\mathcal{I}[y; \theta \mid x]\) derived above.
• Distribution Shift: A Practical Introduction — Why post-hoc Temperature Scaling breaks down when \(p_{\text{test}}(x) \neq p_{\text{train}}(x)\).
• Early-Exit Neural Networks — Using normalized predictive entropy \(\tilde{\mathcal{H}}(x) < \tau_m\) as a dynamic halting criterion.
• Risk-Constrained AI Decision-Making — Converting calibrated predictive probabilities into selective prediction abstention rules.