**Course description:**
This seminar class will focus on new results and directions in machine
learning theory. Machine learning theory concerns questions such as:
What kinds of guarantees can we prove about practical machine learning
methods, and can we design algorithms achieving desired guarantees?
(Why) is Occam's razor a good idea and what does that even mean? What
can we say about the inherent ease or difficulty of different types of
learning problems? Addressing these questions will bring in
connections to probability and statistics, online algorithms, game
theory, computational geometry, and empirical machine learning
research.

The first half of the course will involve the instructor presenting some classic results and background including regret guarantees, combining expert advice, Winnow and Perceptron algorithms, VC-dimension, Rademacher complexity, SVMs, and Kernel functions. The second half will involve student-led discussions of recent papers in areas such as deep learning, multi-task learning, tensor methods, structured prediction, dictionary learning, and other topics

- 08/27: Introduction. The PAC model and Occam's razor.
- 08/29: The Online Mistake-Bound model, Combining Expert Advice / Multiplicative Weights, Regret Minimization and connections to Game Theory.
- 09/03: Sleeping/Shifting experts, the Winnow Algorithm, L_1 margin bounds. [(old) survey article on online learning]
- 09/05: The Perceptron Algorithm, Margins, and intro to Kernels. [more notes]
- 09/10: Do exercises 2,3 and problems 4,5 on this problem set. If you have time, see if you can solve problem 6 (but you don't need to write it up). Also if you have time, look over exercise 1 (but you don't need to write it up). See group-work protocol below.
- 09/12: Do problems 3,4,5 on this problem set. Problem 3 is the hardest. If you have time, do exercise 1 (and think about the total weight). See group-work protocol below.
- 09/17: VC-dimension, Sauer's lemma, and Uniform Convergence bounds. [more notes]
- 09/19: Rademacher bounds: Uniform Convergence II.
- 09/24: Boosting
- 09/26: Kernels, similarity measures, and representations
- 10/01: Semi-supervised learning
- 10/03: Online linear optimization
- 10/08: Bandit algorithms, internal regret, game theory
- 10/10: Work as a group in class on problem set

[2009 version of the course]

**Prerequisites: **A Theory/Algorithms background or a Machine
Learning background.

**Text (recommended):**
*
An Introduction to Computational Learning Theory* by Michael Kearns
and Umesh Vazirani, plus papers and notes for topics not in the book.

**Office hours:** Wed 3-4 or send email to make an appointment.

- [Online learning survey] [Handout on tail inequalities]
- Portions of new book by Shalev-Shwartz and Ben-David (hardcopy only)
- Homework 1. Solutions.
- Homework 2. Solutions.
- Homework 3. Solutions.
- Homework 4. Solutions.
- About projects
- Homework 5. Solutions.

- 01/13: Introduction. PAC model and Occam's razor.
- 01/15: The Mistake-Bound model. Combining expert advice. Connections to info theory.
- 01/20: The Winnow algorithm.
- 01/22: The Perceptron Algorithm, margins, & intro to kernels plus Slides.
- 01/27: Uniform convergence, tail inequalities (Chernoff/Hoeffding), VC-dimension I. [more notes]
- 01/29: VC-dimension II (proofs of main theorems).
- 02/03: Boosting I: Weak to strong learning, Schapire's original method.
- 02/05: Boosting II: Adaboost + connection to WM analysis + L_1 margin bounds
- 02/10: Rademacher bounds and McDiarmid's inequality.
- 02/12: Rademacher bounds II.
- 02/17: MB=>PAC, Support Vector Machines, L_2 margin uniform-convergence bounds.
- 02/19: Margins, kernels, and general similarity functions (L_1 and L_2 connection).
- 02/24: No class today. Nina Balcan talk at 10:00am in GHC 6115.
- 02/26: Learning with noise and the Statistical Query model I.
- 03/03: No class today. Open house.
- 03/05: Statistical Query model II: characterizing weak SQ-learnability.
- 03/17: Fourier-based learning and learning with Membership queries: the KM algorithm.
- 03/19: Fourier spectrum of decision trees and DNF. Also hardness of learning parities with kernels.
- 03/24: Learning Finite State Environments.
- 03/26: MDPs and reinforcement learning.
- 03/31: Maxent and maximum likelihood exponential models; connection to winnow
- 04/02: Offline->online optimization, Kalai-Vempala.
- 04/07: The Adversarial Multi-Armed Bandit Setting.
- 04/09: Game theory I (zero-sum and general-sum games).
- 04/14: Game theory II (achieving low internal/swap regret. Congestion/exact-potential games).
- 04/16: Semi-supervised learning
- 04/21: Some loose ends: Compression bounds, Bshouty's algorithm.
- 04/23: Project presentations
- 04/28: Project presentations
- 04/30: Project presentations

- O. Bousquet, S. Boucheron, and G. Lugosi, Introduction to Statistical Learning Theory.
- N. Cristianini and J. Shawe-Taylor, Kernel Methods for Pattern Analysis, 2004.
- N. Cristianini and J. Shawe-Taylor, An Introduction to Support Vector Machines (and other kernel-based learning methods), 2000.
- Nick Littlestone Learning Quickly when Irrelevant Attributes Abound: A new Linear-threshold Algorithm. This is the Winnow paper from 1987, which also discusses a number of aspects of the online mistake bound model.
- The Adaboost paper: Yoav Freund and Rob Schapire, A decision-theoretic generalization of on-line learning and an application to boosting, Journal of Computer and System Sciences, 55(1):119-139, 1997.
- Robert Williamson, John Shawe-Taylor, Bernhard Scholkopf, Alex Smola Sample Based Generalization Bounds. Gives tighter generalization bounds where instead of using "the maximum number of ways of labeling a set of 2m points" you can use "the number of ways of labeling your actual sample".
- Maria-Florina Balcan, Avrim Blum, and Nathan Srebro Improved Guarantees for Learning via Similarity Functions. Gives formulation and analysis for learning with general similarity functions. Also shows that for any class of large SQ dimension, there cannot be a kernel that has large margin even for all (or even a non-negligible fraction) of the functions.
- Avrim Blum, Merrick Furst, Jeffrey Jackson, Michael Kearns, Yishay Mansour, and Steven Rudich Weakly Learning DNF and Characterizing Statistical Query Learning Using Fourier Analysis. Defines and analyzes SQ dimension. Also weak-learning of DNF via fourier analysis.
- PASCAL video lectures.
- Avrim Blum and Yishay Mansour
Learning, Regret Minimization, and
Equilibria, Chapter 4 of
*Algorithmic Game Theory*, Noam Nisan, Tim Roughgarden, Eva Tardos, and Vijay Vazirani, eds.