Learning Markov Random Fields from data

Monday, April 25, 2016 to Wednesday, April 27, 2016

 

Infotech Oulu Doctoral Program

Lecturer: Dr. Onur Dikmen, University of Helsinki

Credits: 2 ECTS

Date & time:

April 25, 2016 (Monday), 8-16
April 26, 2016 (Tuesday), 8-16
April 27, 2016 (Wednesday), 8-16

Room: IT105

Registration: Contact Burak Turhan (burak.turhan(at)oulu.fi)

This course is an introduction to inference on Markov random fields (MRFs, also known as undirected graphical models), which are widely used models in physics, biology, image processing, and many more disciplines. After providing a brief background on statistical inference (maximum likelihood estimation, Bayesian inference, etc.), a formal definition of MRFs will be given along with some well known examples such as Ising/Potts models and Boltzmann machines. There are two main computational tasks associated with MRFs, both of which prove difficult computationally. The first task (the forward problem) is to calculate statistics given model parameters and is essential in statistical physics. The second task (the inverse problem) is basically the machine learning problem, i.e., using statistical and computational tools to estimate parameters of MRFs from data. Various inference and learning methods (Monte Carlo methods, pseudo-likelihood, variational Bayes, contrastive divergence, score matching, etc.) will be covered to be able to tackle these tasks. The course will end with an obligatory look at the deep learning breakthrough, which is based on MRFs, namely restricted Boltzmann machines. 

Learning objectives

In this course, you will

  • become familiar with influential models from statistical physics
  • know what Markov random fields (MRFs) are
  • be able to come up with your own MRFs!
  • be able to operate on such models and use them to solve practical problems (e.g in image processing)
  • become familiar with powerful methods from statistical inference
  • understand the pros and cons of different inference methods
  • understand the basics of deep learning

About the Lecturer

Dr. Onur Dikmen received his B.Sc., M.Sc. and Ph.D. degrees from the Department of Computer Engineering, Bogaziçi University, Turkey. He worked at Télécom ParisTech, France and Aalto University, Finland as a research associate. He is currently with the Department of Computer Science at University of Helsinki, Finland. His research interests include statistical signal processing, machine learning, and approximate Bayesian inference. He works on inference and optimization in Markov random fields, Bayesian modeling and nonnegative matrix factorization with application to audio source separation and sound event detection.

More information: Burak Turhan

 

Add to calendar

Back to events

Last updated: 1.4.2016