# Theory

> ML Prep practice set.

- Stable ID: `theory`
- Area: machine-learning
- Timebox: 60 minutes

Reason about bias, variance, overfitting, generalization, and evaluation splits.

## Instructions

1. Use observed train-validation patterns rather than vague model-size rules.

## Ordered items

1. [Diagnose model evaluation failures](https://mlprep.iwase.dev/item/mit-6-036-s2019-midterm-q4/?set=theory) — `mit-6-036-s2019-midterm-q4` (medium, 15 min)
2. [Lasso versus ridge penalties](https://mlprep.iwase.dev/item/original-ml-regularization/?set=theory) — `original-ml-regularization` (easy, 6 min)
3. [Read a finite-class generalization bound](https://mlprep.iwase.dev/item/original-ml-finite-class-bound/?set=theory) — `original-ml-finite-class-bound` (medium, 7 min)
4. [Generalization and overfitting knowledge check](https://mlprep.iwase.dev/item/google-mlcc-overfitting-quiz/?set=theory) — `google-mlcc-overfitting-quiz` (easy, 10 min)
5. [BIC for a two-dimensional discrete distribution](https://mlprep.iwase.dev/item/pml-book1-ex4.5/?set=theory) — `pml-book1-ex4.5` (medium, 20 min)

## Completion

Recommend a defensible response to one bias and one variance failure mode.