Integrating Generative Models and Discriminative Models
Time and Place
Mauldin Auditorium (NSH 1305)
There has been a long-time debate about the use of generative models (top-down) and discriminative models (bottom-up) in computer vision and machine learning society. Bottom-up approaches are usually fast once trained but rather static. Top-down approaches are, instead, slow but reflects the perception process of the brain. In this talk, I'll briefly talk about the Data-driven Markov Chain Monte Carlo (DDMCMC) computational paradigm which integrates generative models and discriminative models in a principled way. The efficiency of the bottom-up approaches are decide by how "informative" they are. The DDMCMC paradigm is aimed to tackle a general visual inference problem, "Image Parsing".
I'll then present an algorithm for shape matching and recognition based on a generative model for how one shape can be generated by the other. The matching process is formulated in the EM algorithm. To have a fast algorithm and avoid local minima, we show how the EM algorithm can be approximated by using bottom-up approaches, informative features. The formulation allows us to know when and why approximations can be made and justifies the use of bottom-up features, which are used in a wide range of vision problems. This integrates generative models and feature-based approaches within the EM framework and helps clarifying the relationships between different algorithms for this problem such as shape contexts and softassign.
received his Ph.D. degree in computer science from the
He received the Talbert Abrams award (Honorable mention) by American Society of Photogrammetry and Remote Sensing in 2003. He was also awarded with the David Marr prize at the 9th international conference on computer vision in Nice France.
For appointments, please contact Yanxi Liu.