Selective Prediction

The primary mechanism enabling selective prediction. The model computes a confidence score for each inference. If the score falls below a predefined, domain-specific boundary, the decision is automatically routed to a human review queue rather than being executed.

• Thresholds are calibrated per use case (e.g., 95% for medical diagnosis, 80% for content tagging)
• Prevents low-confidence outputs from reaching end-users
• Directly implements the deferral policy in code

Raw neural network outputs are often poorly calibrated—a softmax probability of 0.9 does not mean the model is 90% likely to be correct. Selective prediction requires post-hoc calibration using techniques like:

• Temperature scaling: Smoothing the softmax distribution on a held-out validation set
• Platt scaling: Fitting a logistic regression model on top of the network's logits
• Isotonic regression: A non-parametric method that corrects systematic overconfidence

Calibrated probabilities are essential for meaningful threshold comparison.

A formal framework where the model is trained with an explicit reject class. Instead of forcing a prediction among K classes, the classifier learns to output "I don't know" when the input lies in an ambiguous region of the feature space.

• Extends standard cross-entropy loss with a rejection cost parameter
• Balances the trade-off between coverage (fraction of inputs predicted) and accuracy
• Commonly used in medical imaging where false negatives carry extreme risk

Selective prediction must identify inputs that differ fundamentally from the training distribution. An OOD detector flags samples where the model has no basis for a reliable prediction.

• Mahalanobis distance: Measures how far a feature representation is from class-conditional Gaussians
• Energy-based models: Use the Helmholtz free energy as a score; OOD samples have higher energy
• ODIN: A preprocessing method that perturbs inputs to amplify the difference between in-distribution and OOD softmax scores

This is critical for models deployed in non-stationary environments.

Training a model to know when it doesn't know requires specialized objective functions. Standard cross-entropy optimizes for accuracy at all costs; abstention-aware losses explicitly penalize confident mistakes.

• SelectiveNet: A deep architecture with an integrated rejection head trained end-to-end with a coverage-constrained loss
• Deep Gamblers: Formulates prediction as a portfolio selection problem where the model can "hedge" by placing a bet on abstention
• Auxiliary abstention heads predict the probability that the primary classifier will err on a given input

The deferral is only as effective as the system that receives it. When a model abstains, the input and all relevant context must be packaged into a structured escalation payload for human review.

• Includes the raw input, model's top-K predictions with confidence scores, and any feature attribution explanations
• Routed to the appropriate expert-in-the-loop based on the predicted class or risk tier
• Logged immutably for audit trail compliance under the EU AI Act

This bridges the gap between model abstention and meaningful human control.