Machine Learning Theory

Fall 2018 (weeks 37 - 51). Lectures on Tuesday 10:15-13:00 at the UvA.

Aim

Machine learning is one of the fastest growing areas of science, with far-reaching applications. In this course we focus on the fundamental ideas, theoretical frameworks, and rich array of mathematical tools and techniques that power machine learning. The course covers the core paradigms and results in machine learning theory with a mix of probability and statistics, combinatorics, information theory, optimization and game theory.

During the course you will learn to

Formalize learning problems in statistical and game-theoretic settings.
Examine the statistical complexity of learning problems using the core notions of complexity.
Analyze the statistical efficiency of learning algorithms.
Master the design of learning strategies using proper regularization.

This course strongly focuses on theory. (Good applied master level courses on machine learning are widely available, for example here, here and here). We will cover statistical learning theory including PAC learning, VC dimension, Rademacher complexity and Boosting, as well as online learning including prediction with expert advice, online convex optimisation and bandits.

MasterMath Website

This course is offered as part of the MasterMath program. To participate for credit, sign up here. We use the MasterMath ELO for submitting homework, forum discussions, and receiving grades.

Prerequisites

The only prerequisites are

Basic probability theory. (in particular conditional probability, expectations, discrete and continuous distributions).
Basic linear algebra (finite dimensional vector spaces, eigen decomposition)
Basic calculus,

all at the bachelor level. The course does require general 'mathematical maturity' though, in particular the ability to combine insights from all three fields when proving theorems.

We offer weekly homework sets whose solution requires constructing proofs. This course will not include any programming or data.

Lecturers

Mode

The grade will be composed as follows.

40%: weekly homework to be handed in before the next lecture.
30%: mid-term exam.
30%: final exam.

It is strongly encouraged to solve and submit your weekly homework in small teams. Exams are personal.

There will be a retake possibility for the exams, which are 60% of the grade.

Schedule

When	What	Who
Tue 11 Sep	Introduction. Statistical learning. Halfspaces. PAC learnability for finite hyp. classes, realizable case. Chapters 1 and 2 in [1].	Rianne
Tue 18 Sep	PAC learnability for finite hyp. classes, agnostic case. Uniform convergence. Chapters 3 and 4 in [1].	Rianne
Tue 25 Sep	Infinite classes. VC Dimension part 1. Chapter 6.1-6.3 in [1].	Rianne
Tue 2 Oct	VC Dimendion part 2. Fundamental theorem of PAC learning. Sauer's Lemma. Chapter 6 in [1].	Rianne
Tue 9 Oct	Proof of Fund.Th of PAC Learning. Nonuniform Learnability, SRM, Other notions of Learning. Chapter 7 in [1].	Rianne
Tue 16 Oct	Rademacher Complexity. Chapters 26 in [1].	Peter
Tue 23 Oct	Midterm exam.
Tue 30 Oct	Full Information Online Learning (Experts). Notes.	Wouter
Tue 6 Nov	Bandits. UCB and EXP3. Chapters 2.2 and 3.1 in [3].	Wouter
Tue 13 Nov	Non-stationary environments. Tracking. Notes.	Wouter
Tue 20 Nov	Online Convex Optimization, Sections 3.1 and 3.3 in [2] and Notes.	Wouter
Tue 27 Nov	Exp-concavity. Online Newton Step. Vovk mixability. Chapter 4 in [2] and Notes.	Wouter
Tue 4 Dec	Log-loss prediction. Normalised Maximum Likelihood. Non-parametric classes.	Peter
Tue 11 Dec	Boosting. AdaBoost. Chapter 10 in [1].	Wouter
Tue 18 Dec	Online learning for solving two-player zero-sum games. Accelerated convex optimisation by means of a reduction to solving games. Notes and this paper.	Wouter
Tue Jan 8	Final exam. 10:00-13:00.

The dependency graph of the second half

Exams

Final

The final exam will test you on the online learning component of the class. Below is the list of topics. The weekly homework prepares you for the exam.

If below we say "algorithm and regret bound" we might question you on the setting, its main algorithm and (do computations involving) the corresponding regret bound. If we add "analysis" you should be able to reprove the regret bound at the exam.

If below we say 'you should read and understand' a section, this means that:

if in the exam we refer to a Lemma/Theorem/Formula in the section, we will always state it explicitly (e.g. on the cheat sheet), so you don't have to learn formulas by heart.
however, you should understand all these Lemmas/Theorems/Formulas since we may ask questions about them and we will not give further explanations of formulas.

Topics

Experts
- Mix loss game, lower bound, Aggregating algorithm regret bound and analysis
- Dot loss game, Hedge algorithm regret bound and analysis
- Vovk mixability. Aggregating algorithm regret bound and analysis.
Bandits
- Stochastic bandit setting. UCB algorithm, regret bound.
- Adversarial bandit setting. EXP3 algorithm, regret bound.
Non-stationarity
- Switching between experts. Fixed Share algorithm. Regret bound and analysis.
Online Convex Optimisation
- Convex losses, Online Gradient Descent regret bound and analysis.
- Strongly convex losses, Online Gradient Descent regret bound and analysis.
- Exp-concave losses, Online Newton Step regret bound.
Log-Loss Prediction:
- Section 9.1-9.7 from Cesa-Bianchi & Lugosi, Prediction, Learning and Games: read and understand completely.
Boosting
- AdaBoost algorithm, iteration bound, analysis.

We will not ask you about VC dimension or PAC learning. That was covered by the mid-term. We will also not ask you about the final lecture about zero-sum games/accelerated optimisation.

Material

We will make use of the following sources

[1] Understanding Machine Learning: From Theory to Algorithms by Shai Shalev-Shwartz and Shai Ben-David.
[2] Introduction to Online Convex Optimization by Elad Hazan.
[3] Regret Analysis of Stochastic and Nonstochastic Multi-armed Bandit Problems by Sébastien Bubeck and Nicolò Cesa-Bianchi.