Notes for the master course Statistical Machine Learning (Prof. Dr. Ulrike von Luxburg, University of Tübingen, Summer Term 2026).

The course does not explain many algorithms. It explains the statistical principles that have to be in place so that machine learning works at all: what the best possible prediction is, when learning from a finite sample converges to it, how the complexity of the function class or the stability of the algorithm controls the gap between training and test error, and why modern overparameterized models still generalize. The last part is about machine learning in society: data and measurement, fairness, explainability and law.

Goal of the course

After this course, students should be able to:

  • state the formal learning setup (risk, loss, Bayes classifier, consistency) and reason inside it
  • derive and compare generalization bounds (finite classes, VC dimension, Rademacher complexity, stability)
  • explain why regularization, bagging and boosting work, and when overparameterized models generalize
  • compute and judge fairness criteria and explanations of real prediction systems

Lectures

#LectureKey concepts
1.1Introduction and Decision Theoryinductive inference, loss and risk, Bayes classifier, regression function, maximum likelihood, MAP, cost-sensitive decisions
1.2Learning from Finite Samplesconsistency, plug-in classifier, empirical risk, ERM, estimation and approximation error, bias-variance decomposition, No Free Lunch
2Perceptron Mistake Boundperceptron as SGD on the linear loss, margin, mistake bound, leave-one-out generalization bound
3Statistical Learning TheoryHoeffding and McDiarmid, uniform convergence, finite classes, shattering coefficient, VC dimension, Sauer-Shelah, Rademacher complexity
4Algorithmic Stabilityuniform stability, stability implies generalization, ERM with strongly convex loss, stability of SGD
5Regularizationleast squares, ridge regression (Tikhonov), stability of regularized ERM, lasso and sparsity
6Aggregation, Bagging and Boostingbootstrap, bagging, random forests, AdaBoost, training and test error bounds, gradient boosting
7Data, Measurement and Validitymeasurement vs. construct, harms in data, features, kernels, random features, benchmarks, four notions of validity
8Overparameterized Learningdouble descent, implicit regularization, minimum norm solution, benign overfitting, why large models
9.1Fairnessdemographic parity, equalized odds, predictive parity, impossibility results, pre-, in- and post-processing
9.2Explainability and Lawfeature importance, SHAP, LIME, counterfactual explanations, EU AI Act, GDPR, copyright

Lecture pages are linked here as soon as they are written.

Terms that appear in several lectures have their own short page under Concepts.

Find a topic

For looking something up later: which lecture and section explains a definition, a theorem or a task type. The table grows with every lecture.

TopicWhere
Inductive bias, why learning needs assumptionsL1.1: Why should ML work at all?
Loss, true risk, Bayes risk, Bayes classifierL1.1: Bayes risk and Bayes classifier
Regression function , Bayes classifier for the 0-1 loss (Theorem 2), Bayes errorL1.1: Explicit Bayes classifier
Hinge, exponential, logistic loss and their Bayes predictorsL1.1: Bayes predictors for various losses
Squared loss: is optimal, Bayes risk L1.1: Regression under the squared loss
ML, MAP, Bayes decision rule with costsL1.1: Bayes decision rule
Exam task: decision boundaries from a density diagram, prior or costs backwardsL1.1: Reading decision boundaries off a diagram
Consistency, universal consistency, Stone’s kNN theoremL1.2: Consistency
Convergence in probability, almost surely, in L1.2: Convergence of random variables
Empirical risk, law of large numbers for a fixed functionL1.2: Proposition 6
ERM and the counterexample where it failsL1.2: Is ERM always successful?
Estimation vs. approximation error, over- and underfittingL1.2: Estimation and approximation error
Classical vs. modern regime, double descentL1.2: Modern regime
Bias-variance decomposition (squared loss)L1.2: Bias-variance
No Free Lunch theoremL1.2: Propositions 7 to 9
Perceptron algorithm, perceptron by hand (exam worksheet)L2: The perceptron by hand
Distance to a hyperplane, marginL2: Margin
Mistake bound (Novikoff) and its proofL2: Theorem 1
Test error of the perceptron, leave-one-out, fast rate L2: Theorem 2
Hoeffding and McDiarmid, error of a fixed function, samplesL3: Error of a fixed function
Uniform convergence, Vapnik-Chervonenkis theoremL3: Uniform convergence
Finite classes, union bound, sample sizesL3: Finite classes
Shattering coefficient, symmetrization, growth function rulesL3: Growth function
Exam task: VC dimension from a growth function tableL3: Growth function tables
VC dimension, Sauer-Shelah, VC boundL3: Sauer-Shelah
VC of linear, margin and neural network classifiersL3: Linear classifiers
Rademacher complexity, summary of all boundsL3: Summary of bounds
Average and uniform stability, generalization gap = average stabilityL4: Proposition 1
Uniform stability bound, which rates work ()L4: Which rates are good enough?
Exam task: stability rate table (yes/no)L4: Stability rates
Strong convexity, ERM is stable with L4: Theorem 3
Smoothness, stability of GD and SGD, early stoppingL4: Stability of GD

How to read these notes

Every lecture page has the same structure: a short TL;DR, the content in the order of the slides, small worked examples, exam-style tasks, a self-test and links to related lectures. The table of contents on the right follows the slide sections. Small labels like Slides 12-15 say where a topic is in the original slides. The notation is the lecturer’s: is the true risk, the empirical risk, the Bayes classifier.

Callouts mark what matters most:

Definition

The exact meaning of a term.

Key formula or theorem

Formulas and statements you should know.

Intuition

Why something works, in plain words.

Common mistake

Typical misunderstandings and wrong exam answers.

Exam relevance

What was asked in the exam or is likely to be asked.

Exam-style task

A task in the format of the real exam: a table or a small data set, then (a) easy, (b) harder, (c) transfer.

Self-test question

Figures marked Interactive are visualizations made for these notes. Move the sliders or click through the steps to see how a method behaves.

Course organization

LecturerProf. Dr. Ulrike von Luxburg
Tutorialsweekly, attendance of at least 9 of 12 tutorials is needed for the exam admission
Assignmentsone sheet per week, handed in by teams of two; at least 50% of the points on average are needed for the exam admission
Gradingthe final exam determines the grade
Exam120 minutes, one A4 page (one side) of handwritten notes allowed
Exam dates28 July 2026 and 8 October 2026
Course pagehttps://www.tml.cs.uni-tuebingen.de/teaching/2026_sml/index.php
  • Exam format: 7 tasks with 4 points each. Task 1 is multiple choice with exactly one correct answer per question. Tasks 2 to 7 have three parts each: (a) an easy computation or reading off a table, (b) a computation plus a check or an interpretation, (c) a transfer question, a construction or reasoning backwards. Most tasks start from a table or a small toy data set. No points are lost for wrong multiple choice answers.
  • The info sheet says that exam tasks are similar to the homework assignments, so the assignment sheets are good practice.

Requirements: linear algebra, multivariate analysis, probability theory, statistics and optimization (for example from Mathematics for Machine Learning), Python, and an introductory machine learning course.

Literature:

  • Moritz Hardt and Benjamin Recht: Patterns, Predictions, and Actions, 2022 (online)
  • Shai Shalev-Shwartz and Shai Ben-David: Understanding Machine Learning: From Theory to Algorithms, 2014 (online)