Talk Topic: Learning Task-Specific Similarity
Gregory Shakhnarovich, Brown University
April 11, 2007

Abstract
The notion of similarity is fundamental in machine learning. The ability to assess similarity, and to find in a database examples similar to a given instance, is central to many statistical learning methods. Similarity is commonly modeled in terms of a distance function in the input space.  However, such a definition may not capture the concept of similarity relevant to the task at hand. In this talk, Shakhnarovich will describe an approach to learning similarity from user-provided examples of what is deemed similar (and, optionally, dissimilar). He will show how to construct, using a greedy algorithm inspired by boosting, an embedding of the data into a weighted binary space, where a simple metric approximates the original similarity concept. This approach provides a means for data compression focused on preserving relevant features, and enables efficient search with respect to user-defined similarity in very large databases. This makes example-based learning work well in some of the problems for which it was previously deemed infeasible.

Bio
Greg Shakhnarovich is a postdoctoral researcher at Brown, where he is working on developing brain-machine interfaces. His current research is focused on machine learning methods for decoding neural signals and using them to control artificial devices such as robotic prostheses. Greg holds a PhD from MIT, where he worked on machine learning for a variety of computer vision problems, and undergraduate and Master's degrees from Hebrew University and the Technion, respectively, in Israel.