Booz Allen Hamilton Colloquium: Gregory Wornell, MIT
Friday, April 8, 2022
3:30 p.m.-4:30 p.m.
Speaker: Gregory Wornell
Department of Electrical Engineering & Computer Science
Massachusetts Institute of Technology
Title: "An Information-Geometric View of Learning in High Dimensions"
We describe a framework for addressing the problem of data feature selection prior to inference task specification, which is central to high-dimensional learning. Introducing natural notions of universality for such problems, we develop a local equivalence among them. We further express the information geometry associated with this analysis, which represents a conceptually and computationally useful learningmethodology. The development will highlight the interrelated roles of the singular value decomposition, Hirschfeld-Gebelein-Renyimaximal correlation, canonical correlation and principle component analyses, Tishby's information bottleneck, Wyner's commoninformation, Ky Fan k-norms, and Brieman and Friedman's alternating conditional expectation algorithm. As will be discussed, thisframework provides a basis for understanding and optimizing aspects of learning systems, including neural network architectures, matrixfactorization methods for collaborative filtering, rank-constrained multivariate linear regression, and semi-supervised learning, amongothers.
Greg Wornell received his PhD from MIT in 1991. Since then, he has been on the faculty at MIT, where he is the Sumitomo Professor ofEngineering in the Department of Electrical Engineering and Computer Science. At MIT he leads the Signals, Information, andThe Electrical and Computer Engineering Distinguished Colloquium Series hosted by Booz Allen Hamilton features distinguished speakers from across the nation and around the globe, and also provides venues in which ECE faculty can showcase their research to a broad audience of their colleagues and students, as well as friends of the university.