CSC 2541 -- Topics in Machine Learning: Natural Scene Statistics

Winter 2005


This course is a graduate seminar devoted to research in natural scene statistics. Recent years have seen a growing interest in the statistics of natural scenes, such as images and sounds. These studies are important on two counts: both for understanding how these statistics influence and are reflected in our perceptual systems, and for developing algorithms that exploit these statistics. On the biological side, a long-standing assumption is that organisms are adapted to the statistical properties of the signals to which they are exposed. Recent developments in statistical modeling, along with powerful computational tools, have enabled researchers to study more sophisticated models for sensory inputs, to validate these models against large data sets, and to begin examining their relevance to representations in neurons.

On the applications front, identifying relevant scene statistics can be useful for solving perceptual inference problems. For example, the distribution of image gradients are different on and off object boundaries, a fact that can be exploited in building effective edge detectors. There are also important applications of statistical knowledge of images to computer graphics, and to the design of devices that interact with humans.

The course will be organized around a set of representative papers in this area. The format of the class will be a mix of background lectures and discussions of these papers. Course requirements include presentation(s), a course project, and class participation. There are no formal pre-requisites. However, some background in linear algebra, probability/statistics, and calculus will be useful. Also, at least one course in AI, preferably in either machine learning and/or vision, would be useful but is not required.

Note that the course differs from previous incarnations of CSC2541, Topics in Machine Learning. The rule for topics courses is that a graduate student can get credit for the same course number more than once, as long as the department can ensure that the courses are quite different and has a subtitle indicating the specific topic.