Statistical learning theory 2022

Материал из Wiki - Факультет компьютерных наук
Перейти к: навигация, поиск

General Information

Lectures: Friday 16h20 -- 17h40, Bruno Bauwens, Maxim Kaledin, room M202 and on zoom

Seminars: Saturday 14h40 -- 16h00, Artur Goldman, room M202 and on zoom (the link will be in telegram)

To discuss the materials, join the telegram group The course is similar to last year.

Colloquium

Rules and questions.

Choose the day: 16 or 17 Dec.

Homeworks

Email to brbauwens-at-gmail.com. Start the subject line with SLT-HW.

Deadline before the start of the lecture, every other lecture.

Sat. 17 Sept 18h10: problems 1.7, 1.8, 2.9, and 2.11
Sat. 01 Oct 18h10: see lists 3 and 4, and 2.10
Fri. 14 Oct 16h20: see problem lists 5 and 6
Sat. 05 Nov 20h00: see problem lists 7 and 8
Sat. 19 Nov 20h00: see problem lists 9 and 10
Sun. 04 Dec 23h59: see problem list 12 (instructions to submit are in telegram)

grades

Course materials

Video Summary Slides Lecture notes Problem list Solutions
Part 1. Online learning
02 Sept Philosophy. The online mistake bound model. The halving and weighted majority algorithms movies sl01 ch00 ch01 list 1 update 05.09 solutions 1
09 Sept The perceptron algorithm. The standard optimal algorithm. sl02 ch02 ch03 list 2 update 25.09 solutions 2
16 Sept Kernels and the kernel perceptron algorithm. Prediction with expert advice. Recap probability theory. sl03 ch04 ch05 list 3 solutions 3
Part 2. Distribution independent risk bounds
23 Sept Sample complexity in the realizable setting, simple examples and bounds using VC-dimension sl04 ch06 list 4 solutions 4
30 Sept Growth functions, VC-dimension and the characterization of sample comlexity with VC-dimensions sl05 ch07 ch08 list 5 solutions 5
07 Oct Risk decomposition and the fundamental theorem of statistical learning theory sl06 ch09 list 6 solutions 6
14 Oct Bounded differences inequality, Rademacher complexity, symmetrization, contraction lemma, quiz sl07 ch10 ch11 list 7 update 15.10 solutions 7
Part 3. Margin risk bounds with applications
21 Oct Simple regression, support vector machines, margin risk bounds, and neural nets sl08 ch12 ch13 list 8 solutions 8
04 Nov Kernels: RKHS, representer theorem, risk bounds sl09 ch14 list 9 solutions 9
11 Nov AdaBoost and the margin hypothesis sl10 ch15 list 10 solutions 10
18 Nov Implicit regularization of stochastic gradient descent in neural nets no seminar
Part 4. Other topics
25 Nov Regression I: fixed design with sub-Gaussian noise notes12 list 12
02 Dec Multiarmed bandids I notes13 list 13
09 Dec Multiarmed bandids II (optional) list 14
16 Dec Colloquium

The lectures in October and November are based on the book: Foundations of machine learning 2nd ed, Mehryar Mohri, Afshin Rostamizadeh, and Ameet Talwalker, 2018. This book can be downloaded from Library Genesis (the link changes sometimes and sometimes vpn is needed).

Problems exam

The exam will happen in the computer room. During the exam
-- You may use handwritten notes, lecture materials from this wiki (either printed or through your PC, Mohri's book
-- You may not search on the internet or interact with other humans (e.g. by phone, forums, etc)


Grading formula

Final grade = 0.35 * [score of homeworks] + 0.35 * [score of colloquium] + 0.3 * [score on the exam] + bonus from quizzes.

All homework questions have the same weight. Each solved extra homework task increases the score of the final exam by 1 point.

There is no rounding except on the final grade. Grades fractional grades above 5/10 are rounded up, those below 5/10 are rounded down.

There are no auto-grades. But because of extra questions and bonuses from quizzes, it might happen that there is no reason to attend the exam. Example: HW = 9/10, 5 extra problems are correct, colloquium=10/10, bonus quizzes = 0.09, exam 0/10. Then the final score is 9.05/10, which is rounded to 10/10.

For students who want to pass with 4/10 with minimal effort: each year I ask to calculate the VC-dimension or Rademacher complexity of some class. It should be easy to have 4/10 for the final exam. If you understand all lecture notes, you pass the colloquium with maximal score. Together this is enough. If only a few students fail and the grades are at least 3.8/10 then failed students may resubmit a few homework tasks to pull up the grade. (This happened in the last 3 years.)


Office hours

Person Monday Tuesday Wednesday Thursday Friday
Bruno Bauwens 15-20h 18-20h
Maxim Kaledin Write in Telegram time is flexible

It is always good to send an email in advance. Questions and feedback are welcome.