Joint Colloquium Distinguished Lecture Series
Visual Learning and Multimodal Interaction
Wednesday, September 19, 2007
306 Soda Hall (HP Auditorium)
4:00 - 5:00 pm
Vision Interface Group, Computer Science and Artificial Intelligence Laboratory (CSAIL), Massachusetts Institute of Technology
Intelligent systems should support natural interaction with people, and be aware of and reactive to objects in the environment. Computer vision plays a key role towards both goals, enabling multimodal conversation and recognition of objects and object categories. In this talk I'll present new work on the recognition of visual agreement and grounding gestures in conversation, to allow conversational systems to directly perceive visual backchannel cues. Our work exploits dialog context as part of the visual recognition process, leading to improved robustness. I'll also overview our work on efficient algorithms for object category recognition and indexing, including linear time kernels which can compute approximate partial match correspondences over sets of local image features, and a sub-linear time hashing scheme for this representation. A key challenge in perception is the question of how to discover good representations for learning, and I'll close by describing new transfer learning methods for visual recognition, in which representations learned from previous tasks speed learning of future tasks. I will show results demonstrating that transfer of both linear and non-linear representations improves performance on visual recognition tasks.
Trevor Darrell leads the Vision Interface Group at the MIT Computer Science and Artificial Intelligence Laboratory. His interests include computer vision, multimodal interfaces, and machine learning. Previously, he worked as a Member of the Research Staff at Interval Research in Palo Alto, CA. He received his PhD and SM from MIT in 1996 and 1991, respectively, while working at the Media Laboratory, and the BSE from the University of Pennsylvania in 1988, where he worked in the GRASP Robotics Laboratory.
|Return to EECS Joint Colloquium|