Module 01: Introduction
Course framing module: what statistical learning is, the two organizing axes (supervised/unsupervised, prediction/inference), and the three umbrella problem types (regression, classification, unsupervised). One lecture (Jan 5), no recommended exercises tied to this module, no compulsory-exercise problems. Load-bearing for the exam: the Q1-style fill-in-the-blank that uses this module’s vocabulary to tag a real-world scenario, flagged by the prof in L27-summary.
Lectures
- L01-intro: course mechanics, the prof’s framing of statistical learning vs ML / classical stats / data science, three worked-example problem types (Framingham regression, Fisher iris LDA, rat-gene-expression hierarchical clustering)
Concepts (atoms in this module)
- statistical-learning: prof’s framing of the field; “statisticians getting in on machine learning”; misspecified-model regime; the two orthogonal axes that organize every method in the course
- prediction-vs-inference: same model, two uses; finance vs science (Bitcoin forecasting vs Framingham SBP); same data → very different design choices
- supervised-vs-unsupervised: first organizing axis: have a Y to aim at vs not; LLMs as “supervised in disguise”; pure unsupervised flagged as “dangerous statistics” without a downstream check
Cross-cutting concepts touched (Specials)
- bias-variance-tradeoff: first introduced module 02; previewed here in L01-intro as the statistician’s-angle contribution to ML
- cross-validation: first taught module 05; previewed here as the supervised-side antidote to unsupervised’s “how do I know this works?”
Exercises
The recommended-exercise series starts at exercises/Exercise2/ (module 2). Module 1 has no recommended exercises and no compulsory-exercise problems. The conceptual material is tested via Q1-style fill-in-the-blank using the vocabulary of supervised-vs-unsupervised and prediction-vs-inference. The closest drill problem lives in module 02:
- Exercise2: problem 2.1 (“describe a real classification and a real regression task, identify response/predictors, decide prediction or inference”) is the canonical drill for this module’s vocabulary
Out of scope (this module)
- Proper-noun history of the field (McCulloch & Pitts, Rosenblatt, AlexNet, Bell Labs checkers bot, dates and inventors) - “I’m not going to ask you a history question on the test” - L22-unsupervised-2 / L27-summary. The framing matters; the names and dates do not.
- Etymological / definitional questions on the field labels (“define machine learning,” “define AI”) - “I won’t ask any questions about how they’re defined” - L02-statlearn-1. The prof flags ML / AI / statistical learning / data science as branding more than substance.
- R / Python package names, function syntax, executable code - L27-summary: “no language, no memorizing package names, no language-specific coding.”
- Course-mechanics trivia (compulsory-exercise pass marks, group sizes, Rmd vs Python policy, who taught it last year): administrative only; not exam content.
- Data-science pipeline steps 1–3 (hypothesis formulation, scraping, structuring) - L01-intro: explicitly out of scope; statistical learning owns steps 4–6 (modelling, fitting, communication).
ISLP pointer
Chapter 1: Introduction. Mostly motivational (Wage / Smarket / NCI60 examples, brief field history). Atoms carry section-level isl-ref: pointers: prediction-vs-inference → §2.1.1–§2.1.2, supervised-vs-unsupervised → §2.1.4. The historical rundown in §1 (least squares 1800s, LDA 1936, GLMs 1970s, trees 1980s, NNs, SVMs) is explicitly out of scope per the no-history-questions rule.