A framework that studies the conditions under which algorithms can generalise from finite training data to unseen data, providing theoretical bounds on prediction error via concepts such as VC dimension, PAC learnability, and Rademacher complexity.
Semantic Classification
Content
- Statistical learning theory analyses how well a model fitted on a sample will perform on the underlying distribution. Concepts such as the bias-variance trade-off, capacity measures, and generalisation bounds explain why over-flexible models overfit and how regularisation controls this.
- Its results provide the formal justification for empirical risk minimisation and guide model selection across supervised learning methods.