Phd Prospectus Defense - Local Learning by Partitioning Enviroment: Joseph Wang

Our goal is to exploit local structure for various learning problems. We assume that data is \well-behaved" locally such that learning can be accomplished accurately in a neighborhood, using simple functions locally whereas complex functions are required globally, or using few sensors locally as opposed to the full set globally. We focus on the problems of learning local neighborhoods based on non-linear structures, approximating complex decision boundaries by piecewise locally linear functions, and learning to adaptively acquire sensors locally to reduce the number of sensor measurements during test-time.