Generalization
An AI model’s ability to perform well on new, unseen data by capturing underlying patterns rather than memorizing training examples.
Definition
The core property that distinguishes effective models from overfit ones. Generalization is achieved via appropriate model capacity, regularization techniques (dropout, weight decay), data augmentation, and robust validation (cross-validation, hold-out sets). Governance involves monitoring generalization gaps (train vs. validation error), setting acceptable thresholds, and retraining models when performance on production data diverges significantly from test results.
Real-World Example
An image-classification team observes that their model’s training accuracy is 99% but test accuracy is 75%. They introduce data augmentation (rotations, color jitter), apply dropout layers, and retrain, achieving balanced train/test accuracies around 90%, demonstrating improved generalization before deployment.
.svg)
