Statistical Learning Theory

by Bruce Hajek and Maxim Raginsky

This page contains lecture notes for ECE 543, Statistical Learning Theory:

Revision log:

Mar 18, 2021
minor typos fixed in Chapter 8
Mar 11, 2021
a key monotonicity condition is added to Theorem 8.4
Feb 9, 2021
minor typos fixed in Chapters 2 and 5
Jan 28, 2021
added a discussion of interpolation without sacrificing statistical optimality (Section 1.3)
May 16, 2019
too many changes to catalog, lots of bugfixes and new content
Apr 4, 2018
added a section on the analysis of stochastic gradient descent (Section 11.6)
added a new chapter on online optimization algorithms (Chapter 12)
Mar 28, 2018
revised and extended the section on convex analysis in Hilbert spaces (Section 11.1)
revised the section on stochastic gradient descent (Section 11.5)
Mar 27, 2018
Rademacher complexity bounds for neural nets (Section 6.5)
Feb 28, 2018
revised basic bounds via surrogate losses (Section 6.2)
added a section on AdaBoost (Section 6.4)
Feb 19, 2018
revised section on structural results for Rademacher averages (Chapter 4)
statement and proof of the contraction principle (Chapter 4)
proof of the Sauer-Shelah lemma via Pajor's theorem in Chapter 5
Feb 7, 2018
added learnability of finite concept classes in the model-free framework
streamlined the symmetrization argument in Chapter 4
streamlined the proof of Finite Class Lemma in Chapter 4
Jan 25, 2018
added PAC learnability of finite concept classes
Jan 18, 2018
streamlined the proof of McDiarmid's inequality

Note to instructors: You are welcome to use any portion of these lecture notes in your own classes without asking our permission, but please give us proper credit and please include a link to this web page (http://maxim.ece.illinois.edu/teaching/SLT) for the most recent revision.

Feedback is always welcome, especially bug reports. We particularly appreciate hearing from students or instructors outside UIUC who find this stuff useful (or useless).

Creative Commons License
Statistical Learning Theory by Bruce Hajek and Maxim Raginsky is licensed under a
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.