- Block 9 — Generalization
- Objective
- state and prove the decomposition of expected squared error
into squared bias, variance and irreducible noise; define each term without
ambiguity and measure all three by simulation rather than assert them;
recognize the diagnostic signature of excess bias and that of excess variance;
know which lever acts on which component and by how much, with measured
numbers for regularization, for bagging, for boosting and for training-set
size; understand why the identity is exact under squared loss and only
approximate — or sign-ambiguous — under 0-1 loss; and know the boundary the
overparameterized regime imposes on the framework.
- Estimated duration
- 75 minutes
- Prerequisites
- chapters 026 (train/validation/test split), 030 (error,
loss and metric) and 031 (overfitting and underfitting)
- Associated quizzes
032.1-quiz-the-decomposition.md to
032.8-quiz-double-descent.md