T-61.183 Special course in Information Science III (4 ov) (L)

Spring 2003

Place: Seminar room T4 in T-building
Time: Mondays 14 - 16, starting on January 20
Language: English
Credit points: 4
Lecturer: Prof Juha Karhunen
Assistant: M.Sc. Karthikesh Raju
Web page: http://www.cis.hut.fi/Opinnot/T-61.183/

Support Vector Machines and Kernel Methods [pdf]

Support vector machines and related kernel methods are currently a hot topic in machine learning. They are based on the idea of an implicitly computed feature mapping which allows one to implement 'nonlinear' classifiers and representations using linear techniques. Several factors have contributed to the current popularity of kernel methods. They are theoretically much more tractable than for example neural networks; their statistical performance is usually very good; and there are available a large number of efficient algorithms and implementations.

Materials

B. Schölkopf and A. Smola: Learning with Kernels - Support Vector Machines, Regularization, Optimization, and Beyond, The MIT Press, 2002, xviii+626 pages.

The book has been received quite well, and more information on it can be found on its home page

This page also provides Chapters 1,2,3,5,6,7, Appendix A in pdf format. Useful auxiliary material complementing the book are also available, including other related books, tutorial articles and talks, and software at another related webpage

Prerequisites

The course is intended mainly for graduate students, but it can be taken also by undergraduate students who are mathematically mature enough, having already passed most of the studies required for the Dipl.Eng. degree. Even though the mathematics required in the remainder of the book are presented in the first part of the book, a basic knowledge of probability theory and linear algebra is necessary. Some knowledge of pattern recognition and/or neural networks is also useful but not necessary. The seminar would be in english.

Requirements for passing the course

To pass the course with 4 credits, you have to at least

• Participate Sufficently in the seminar meetings
• Give one's own talk(s)
• Solve a sufficient percentage of problems choosen from the book
• Perform the given computer assignment(s)

Timetable

The first meeting will be on Monday 20th January at 14:15 in the lecture room T4 in the Computer Science Building at HUT. Note that the room has changed (from the earlier announced TB353). After this, the course continues weekly at the same time in the same place. The details of the course will be determined later on after the number of the participants is roughly known.

Seminar Program and Timetable.

Contacts

The responsible teacher of the course is Prof. Juha Karhunen, email Juha.Karhunen@hut.fi, room TB327, tel. 451 3270. You can contact him during the seminar meetings or otherwise preferably at reception time, on Fridays 15-16. The course assistant is MSc. Karthikesh Raju, email: karthik@james.hut.fi, room TB330.

Talks

• Introduction - Prof. Juha Karhunen [Prof. K-R.Müeller] - 2003.01.20
• Kernels - Karthikesh Raju - 2003.02.03
• Risk and Loss Functions - Jan Eriksson - 2003.02.10
• Regularization - Alexandre Iline - 2003.02.17
• Optimization - Tapani Raiko - 2003.03.03
• Pattern Recognition [7.1 - 7.4] - Ramunas Girdziusas - 2003.03.10
• Support Vector Regression - Pasi Lehtimäki- 2003.03.17
• SVC Extensions - Juha Raitio - 2003.03.17
• Kernel PCA - Anne Patrikainen - 2003.04.07

Exercise Problems

• Talk 2 : 2.1, 2.33, 2.35
• Talk 3 : 3.1, 3.3, 3.13
• Talk 4 : 4.2, 4.3
• Talk 5 : 5.2, 5.5
• Talk 6 : 6.4 (contains several subproblems)
• Talk 7 : Computer Asignment - as defined here
• Talk 8 : 7.9 ,7.11
• Talk 9 : 8.2, 8.10
• Talk 10 : 9.1 ,9.2
• Talk 11 : 14.1, 14.3
• Talk 12 : 15.1 ,15.9

Welcome

Prof. Juha Karhunen