COMP 875: Machine Learning Techniques in Image Analysis

Fall 2009, Tuesdays and Thursdays 3:30-4:45, FB 009

Instructor: Svetlana Lazebnik  (lazebnik -at- cs.unc.edu)

Quick links: schedule, reading materials, useful resources

About the Course

This is a graduate seminar course on statistical and machine learning techniques. It is intended as a standard part of the medical imaging and computer vision graduate curriculum, to be taken by students who have already taken COMP 665, 775, or 776. However, students from other research areas who want to get a better background in machine learning are also encouraged to attend. For example, students who have taken a Data Mining course would be well prepared for this one.

The course will consist of a series of introductory lectures, followed by student presentations. The selection of presentation topics will be driven by the interests of students who are enrolled. Possible topics include, but are not limited to, the following:

• Classifiers: linear models, support vector machines, boosting
• Kernel methods
• Bayesian methods, EM
• Random field models
• Sampling techniques such as Markov Chain Monte Carlo
• Unsupervised learning: density estimation, clustering
• Manifold learning and dimensionality reduction
• Distance metric learning
• Semi-supervised learning
• Online and active learning
• Sequential inference (i.e., tracking)
• Large-scale learning

Presentations

• Integration: think of yourself as professor for a day. You should strive to give a comprehensive and understandable lecture introducing the class to a specific topic in machine learning. Even if your presentation centers on one particular method or application, you need to consult multiple sources and give a survey of relevant literature.
  
  
• Critical thinking: you are expected to give a balanced and critical perspective on the topic you are presenting. You should go beyond simply giving a list of relevant papers or describing the techniques. Compare and contrast different papers, question assumptions, expose possible flaws in experimental protocol, identify limitations, suggest alternative applications and/or directions for future research.
  
  
• Interactivity: try to involve the class. Ask for input from the class, try to think of discussion topics. This part should follow naturally if you get in the "critical thinking" mindset.

Project

• Implementation: do an empirical evaluation of one or more methods, apply an existing method to an interesting problem, extend an existing method, or design and implement your own solution to a learning problem.
  
  
• Survey paper: you can think of this as an extended set of notes produced based on your presentation slides. It should serve as a thorough introduction for anyone who wishes to learn about your topic. The survey should be structured as a formal academic paper. It should be 10-15 pages in length (single-spaced, single colum, 11pt) and typeset in LaTeX.

Grading

• Participation and attendance: 30%
• Presentation (including reading list, preliminary slides and meeting, and final slides): 35%
• Project (including proposal, progress report, and final report): 35%

• NETLAB --- MATLAB software to go with C. Bishop's book
• UCI Machine Learning Repository --- source of datasets for ML problems
• List of object recognition datasets --- compiled by Kristen Grauman
• Machine Learning Summer School/Workshop --- University of Chicago, summer 2009
• Face Recognition Homepage