6.892 Statistical Computer Vision and Learning Theory

or, how to train your brain.


When most creatures are born they cannot see. Between birth and adulthood some set of changes occur that enable vision. Ever wonder what an infant's brain might be learning as observes the world? Is there a simple theory that might explain the computations that underlie all of vision?

This course will cover a branch of computer vision and learning that is based on probabilistic analysis. Our primary focus will be on images, but we will be covering a number of more general subjects such as Bayesian decision theory, unsupervised learning, simulated annealing, and markov random fields (hidden and not hidden). In vision we will cover theories of unsupervised learning such as the Helmholtz(TM) machine, and Kohonen nets. We will ask and even try to answer the question, "What are edges for?" We will talk about supervised learning techniques such as neural networks, and radial basis functions

Some theoretical knowledge of linear algebra and probability is a must. Interest in the brain a benefit. Some basic knowledge of computer vision and graphics will be assumed.


Professor Paul A. Viola
viola@ai.mit.edu (preferred point of contact)
Room NE43-733 Phone x3-8828
Check out the new bibliography.
Tuesday and Thursday 1:00 to 2:30 PM
Class will consist both of lectures and guided discussion of papers. Students are strongly encouraged both to read and form opinions about the papers presented. Discussion based on theoretical or empirical analysis is especially encouraged.
Lectures (Hey it works now!!)
  • There is one required project or paper which must be presented in class.
  • A number of Matlab based example labs will be made available to the class. Solutions to these will not be graded but can form the basis of a class project.
  • Text
    A collection of papers from the literature. Sections from
    Recommeded Prerequisites