Modern Approaches to Machine Learning
(Winter term; 2h of lectures per week)
This course addresses graduate students in Computer Science
during the initial period of their PhD.
See also the page of the
Graduate School in Computer Science
Adaptive methods to programming and to `intelligence'
are of increasing importance in Computer Science.
In this course some important approaches are discussed.
Specifically, the course will concentrate on modern concepts
of adaptive intelligence that arise in the context
of supervised learning and machine learning.
Most of these methods have been inspired by
research in neural networks,
but are now recognized as much more general principles
with a wide area of application.
The course in Winter 2002/2003 covers the following topics.
Part I. Lectures
- 22.10. Introduction: pattern recognition and classification; simple perceptrons
- 29.10. Artificial Neural Networks: Multilayer Perceptrons and BackProp
- 5.11. Generalization and Regularization
- 12.11 Classical statistical approaches to classification
- 19.11. Maximum likelihood, mixture models, and Expectation Maximization (EM)
- 26. 11. Support Vector Machines (SVM)
- 3.12. SVM: Quadratic Programming, Optimization under constraint
- 10. 12 Comparison of supervised approaches (Radial basis functions, Gaussian mixture models and Fuzzy Logic).
Introduction to Reinforcement Learning
- 17.12. Reinforcement Learning: Bellman equation, Q-learning,
- 17.12. Reinforcement Learning: Q-learning in continuous space and time
- Part II. Miniproject
Comparison of 3 different approaches (EM, BackProp, SVM) on a
- C. Bishop, Neural Networks and Pattern recognition,
- S.Haykin, Neural Networks, Prentice Hall, 1994
- R.O. Duda and P.E. Hart and D.G. Stock, Pattern Classification,
John Wiley, 2001
N. Scholkopf and A.J. Smola, Learning with kernels:
support vector machines, regularization, optimization, and beyond, MIT press,
- Sutton and Barto, Reinforcement Learning, MIT Press.
LCN Home Page
Home Page I&C