På Svenska

EN3202 Pattern Classification and Machine Learning 8.0 credits

Mönsterigenkänning och maskininlärning

The course aims to provide a comprehensive theoretical treatment of probabilistic approaches in pattern recognition and machine learning, with emphasis on Bayesian learning.

After passing this course the student should be able to

  • describe the general principles of probabilistic pattern classification and Bayesian parameter estimation,

  • analyze previously unsolved problems in data classification or regression, for example problems encountered in the student's own research project, and formulate a theoretical probabilistic model for the selected application,

  • apply Bayesian parameter estimation, by selecting, if necessary, a suitable approximation approach to make the problem computationally tractable,

  • understand and critially analyze new probabilistic pattern-recognition and machine-learning methods proposed in the scientific literature by other researchers.

  • Educational level

    Third cycle

  • Academic level (A-D)

    D

  • Subject area

  • Grade scale

Information for research students about course offerings

The course is given in the period March - June, including about 10 weekly seminars, each 2-3 hours, and, finally, an individual 72-hour open-book exam.

Learning outcomes

After passing this course the student should be able to

  • describe the general principles of probabilistic pattern classification and Bayesian parameter estimation,

  • analyze previously unsolved problems in data classification or regression, for example problems encountered in the student's own research project, and formulate a theoretical probabilistic model,

  • apply Bayesian parameter estimation, by selecting, if necessary, a suitable approximation approach to make the problem computationally tractable,

  • understand and critially analyze new probabilistic pattern-recognition and machine-learning methods proposed in the scientific literature by other researchers

Course main content

After an initial review of probabilistic models for multivariate data and the principles of Bayesian learning as opposed to point estimates of model parameters, these models and methods are further developed for various applications in regression and classification. The following main topics are covered:

  • Generalized linear models for regression and classification

  • Neural networks

  • Kernel methods, especially sparse approaches, such as the Relevance Vector Machine (RVM) and Support Vector Machine (SVM)

  • Graphical models, incl. Bayesian networks and Markov random fields

  • Mixture models and Expectation Maximization

  • Approximate inference methods, e.g., variational inference with factorized approximation

  • Monte Carlo sampling methods

  • Probabilistic (Bayesian) principal component analysis (PCA)

  • Models for sequencial data, especially Hidden Markov models

Disposition

About 10 weekly seminars, 2-3 hours each. In each session the discussion is focused on one main topic selected from the course book. Students demonstrate and discuss solutions to selected exercise problems.

Eligibility

PhD students in Electrical Engineering or Computer Science or advanced MSc candidates with solid background in probability theory. The undergraduate course EN2202 Pattern Recognition is a recommended but not compulsory prerequisite.

Prerequisites

The course is intended mainly for PhD students in Electrical Engineering or Computer Science.

Literature

Bishop, C.M (2006). Pattern recognition and machine learning. Springer.

Required equipment

Computer with Matlab.

Examination

Examination is based on active participation in course seminars and a final written individual open-book exam.

Requirements for final grade

Active participation is required in at least 70% of course meetings. Individual 72-hour open-book exam with given advanced problems, of which at least 50% must be correctly solved.

Offered by

EES/Sound and Image Processing

Contact

Arne Leijon

Examiner

Arne Leijon

Add-on studies

A solid background in Probability Theory is required. Basic knowledge corresponding to the undergraduate course EN2202 Pattern Recognition is recommended.

Version

Course plan valid from: Spring 11.