SIGGRAPH 2004 course: Introduction to Bayesian Learning

SIGGRAPH course page

Sophisticated computer graphics applications require complex models of appearance, motion, natural phenomena, and even artistic style. Such models are often difficult or impossible to design by hand. Recent research demonstrates that, instead, we can "learn" a dynamical and/or appearance model from captured data, and then synthesize realistic new data from the model. For example, we can capture the motions of a human actor and then generate new motions as they might be performed by that actor. Bayesian reasoning is a fundamental tool of machine learning and statistics, and it provides powerful tools for solving otherwise-difficult problems of learning about the world from data. Beginning from first principles, this course develops the general methodologies for designing learning algorithms and describes their application to several problems in graphics.

Organizer and Lecturer
Aaron Hertzmann
University of Toronto

Course notes

Course slides

Related Textbooks



3:45pm-4:00pm: Introduction

  • The future of graphics: data-driven analysis and synthesis
  • The need for Bayesian reasoning

    4:00pm-4:45pm: Fundamentals of Bayesian probabilistic reasoning

  • Classical (Aristotelian) logic and its limitations
  • Cox axioms
  • Bayes Rule
  • Parameter estimation vs. Bayesian prediction
  • Bayesian prediction
  • Learning multinomials and Gaussians
  • Overfitting and underfitting
  • Regression
  • Applications in graphics

    4:45pm-5:15pm: Statistical shape and appearance models

  • Principal Components Analysis (PCA)
  • Probabilistic PCA
  • Applications in graphics and vision: faces and bodies

    5:15pm-5:30pm: Summary and Conclusions

  • Pros and cons of the Bayesian approach
  • Audience questions