Seminars/Colloquia/Invited Talks

Seminars

Zhuowen Tu

On the Relationship between Generative and Discriminative Models

PLACE:	Clark 314
EVENT:	CIS Seminar
DATE:	April 3, 2007
TIME:	1:00 - 2:00 PM

Abstract

Learning effective and efficient generative and discriminative models is a central task in computer vision and medical imaging. Both the models are widely used in a variety of problems such as classification, detection, and segmentation.
In the first half of my talk, I will present a hybrid approach which combines the generative and discriminative models directly in the modeling stage. The discriminative model selects and fuses a set of cues (features) from a large number candidate pool to implicitly account for local photometric and geometric information. The generative explicitly takes global shape information into account. I will show two applications using the hybrid models, namely brain sub-cortical segmentation and major sulci detection.
In the second half of the talk, I will present a new framework which progressively learns a target generative distribution through discriminative approaches. From the generative model side: (1) A reference distribution is used to assist the learning process, which removes the need for sampling processes in the early stages. (2) The classification power of discriminative approaches, e.g. boosting and SVM, is directly utilized. (3) The ability of selecting/exploring features from a large candidate pool allows us to make nearly no assumptions about the training data. From the discriminative model side: (1) This framework improves the modeling capability of discriminative models. (2) It can start with source training data only and gradually "invent" negative samples. (3) Sampling schemes can be introduced to discriminative models. (4) The learning procedure helps to tighten the decision boundaries for classification, and therefore, improves robustness. I will show a variety of applications such as texture modeling/classification, image analogies, face modeling/detection, natural image statistics learning/image de-noising.

Brief Biography:

Zhuowen Tu received his Ph.D. in computer science from Ohio State University in 2002. He was a post-doc in the statistics department at UCLA from 2002 to 2004. He then went to Siemens Corporate Research as a member of technical staff. He is now a researcher in the lab of neuro imaging, UCLA. Zhuowen Tu (with collaborators) received David Marr prize in ICCV 2003 and he is a recipient of the Talbert Abrams award (Honorable Mention) from American Society of Photogrammetry and Remote Sensing.