Uncertainty Quantification

A class of neural networks that place a probability distribution over model weights rather than learning point estimates. By performing inference over these distributions, BNNs naturally produce predictive uncertainty. Key approaches include:

• Variational Inference: Approximates the true posterior with a simpler distribution.
• Monte Carlo Dropout: A practical approximation that uses dropout at test time to sample from the weight posterior.
• Epistemic Uncertainty: Captures model ignorance due to lack of data, which is reducible with more training examples.

A fundamental decomposition of predictive uncertainty into two distinct types:

• Aleatoric Uncertainty: The irreducible noise inherent in the data itself, such as ambiguous cell boundaries or overlapping tissue features. It is homoscedastic (constant) or heteroscedastic (input-dependent).
• Epistemic Uncertainty: The model's ignorance due to limited knowledge or data. This is reducible by collecting more representative training samples from diverse scanners and staining protocols. Distinguishing between them is critical for triaging: high epistemic cases need a second opinion, while high aleatoric cases are inherently ambiguous.

A distribution-free, model-agnostic framework that produces prediction sets with a formal finite-sample coverage guarantee. Instead of a single label, it outputs a set of plausible classes that contains the true label with a user-specified probability (e.g., 95%).

• Validity: The coverage guarantee holds under the minimal assumption of exchangeability.
• Adaptive Prediction Sets: Set size reflects difficulty—a clear Gleason 3 pattern yields a small set, while an ambiguous pattern yields a larger set.
• Clinical Utility: Provides a rigorous, interpretable alternative to softmax probabilities for triaging difficult cases.

A simple yet highly effective method for uncertainty estimation that trains multiple models with different random initializations and aggregates their predictions. Despite its simplicity, it remains a strong baseline.

• Diversity: Each ensemble member converges to a different local minimum, capturing distinct functional modes.
• Variance as Uncertainty: The disagreement between ensemble members quantifies epistemic uncertainty.
• Robustness: Ensembles also improve accuracy and calibration, making them a practical choice for production diagnostic systems where computational budget allows.

A scalar metric that quantifies the miscalibration of a model's confidence estimates. It measures the gap between predicted confidence and empirical accuracy.

• Calculation: Predictions are binned by confidence; ECE is the weighted average of the absolute difference between accuracy and confidence in each bin.
• Reliability Diagrams: Visualize calibration by plotting accuracy against confidence.
• Clinical Impact: A well-calibrated model that says it is 90% confident is actually correct 90% of the time. Overconfident misdiagnoses erode clinical trust and pose patient safety risks.

An inference-time technique that applies random data augmentations (e.g., rotations, flips, color jitter) to a single input and aggregates the predictions. The variance across augmented predictions serves as an uncertainty estimate.

• Mechanism: Each augmented view is a slightly different perspective; consistent predictions indicate confidence, while divergent predictions signal ambiguity.
• Practicality: Requires no model modification and works with any pre-trained classifier.
• Limitation: Primarily captures aleatoric uncertainty related to input transformations, not full epistemic uncertainty about the model's knowledge.