M12: Summary and Exam Review

Module 12: Summary and Exam Review

Meta-router. Module 12 has no atoms of its own, it is the exam-review session, and all exam-relevant content lives in atoms owned by earlier modules. This MOC routes to the canonical exam-review lecture, the scope authority, and the cross-cutting Specials any module-12 query will pass through.

Lectures

• L27-summary: Apr 28 dedicated exam Q&A: logistics, the scope rule (verbatim), problem-by-problem walkthrough of the 2025 paper showing how each question is reformatted for the 2026 open-book exam, plus the “mathy question” template (MLE = LS under Gaussian errors)

Scope authority

• scope: the canonical “is X in scope?” reference: source hierarchy (exercises > lectures > slides; ISLP for fleshing out, not for scoping), explicit exclusions with verbatim anchors, programming policy, 2026 question patterns, past-exam translation table, exam logistics

Cross-cutting concepts touched (Specials)

These are the six cross-module atoms that any module-12 query will likely route through. Each is owned by no single module, threading through the whole course.

• bias-variance-tradeoff: promised exam question per L27-summary (“definitely going to be a question”); the derivation (CE1 problem 1b) is the canonical “mathy” template
• regularization: the prof’s “central trick of statistical learning”; ridge / lasso / weight decay / dropout / smoothing-spline λ / tree pruning α / implicit SGD all unify here
• cross-validation: the prof’s preferred hyperparameter selector everywhere; right-vs-wrong-way is a flagged trap
• standardization: mandatory pre-processing before ridge, lasso, PCA, PCR, k-means, hierarchical, KNN, NNs; one-line diagnosis for “results look weird”
• multivariate-normal: foundation for OLS sampling distribution and LDA/QDA class-conditionals; CE1 problem 1g matches contour plots to Σ
• double-descent: prof’s hobbyhorse; explains why “tradeoff” framing is incomplete; recurs in L04, L11, L13, L24, L26

Highest-leverage atoms for exam prep

Drawn from L27-summary’s walkthrough plus the signals catalogued in scope. These are the atoms most likely to anchor an exam question:

• least-squares-and-mle: the L27-flagged “mathy” template (Gaussian → MLE = LS)
• ridge-regression, lasso: Q3a, Q6b on the 2025 paper; bias-variance reasoning applied
• logistic-regression, odds-and-log-odds: Q3c (compute odds ↔ probability) + Q7 (interaction trap on odds-ratio)
• linear-discriminant-analysis, discriminant-score-and-decision-boundary: “this would be a typical exam question,” said twice in L09
• categorical-encoding-and-interactions: Q2 and Q7 interaction trap (main-effect coefficient is local to interacting variable = 0)
• smoothing-splines: Q3d direction-of-effect trap (higher λ → smoother, not wigglier)
• principal-component-analysis, explained-variance-and-scree-plot: Q3e cumulative-PVE hand calculation
• hierarchical-clustering: Q5 by-hand dendrogram with complete linkage (“−1 point per mistake” warning)
• nn-parameter-count: Q3b explicit prof-flagged calculation (don’t forget biases)
• sensitivity-specificity, roc-auc, confusion-matrix: Q7 cont; write the formula even when not plugging in numbers
• random-forest, boosting, gradient-boosting: Q6c, Q7 cont; tune hyperparameters by CV, interpret why they beat linear models

Out of scope

This module inherits the whole course’s exclusion list. See scope §“Explicit out-of-scope” for the full catalogue with verbatim anchors. Highlights include: SVM (entire ISLP ch. 9), survival analysis, multiple-testing corrections, AIC/BIC/Cp derivations, F-test mechanics, Bayesian-prior interpretation of ridge/lasso, natural-spline basis math, detailed boosting pseudocode, advanced NN internals (skip connections, Adam, BatchNorm, LSTM/GRU gates, universal-approximation proof), and all R/Python package names and executable code.

ISLP pointer

No single chapter, as module 12 is integrative. For deep treatment of any in-scope topic, route to the chapter owned by the relevant earlier module. ISLP is open-book on the exam; the A5 sheet should hold what ISLP’s index doesn’t get you to fast (interaction-trap reminders, direction-of-effect cheats, parameter-count formulas).