Dartmouth Events

Inference in Structured, Combinatorial and Complex Domains

Siamak Ravanbakhsh: probabilistic inference gives a generic recipe to many modern applications, where data is often abundant, unlabeled and structured.

Wednesday, March 1, 2017
4:30pm – 5:30pm
Carson L02
Intended Audience(s): Public
Categories: Lectures & Seminars

Abstract:   In this talk I will argue that probabilistic inference gives a generic recipe to many modern applications, where data is often abundant, unlabeled and structured. In particular, structured data has an ever-growing presence in applications, from high-throughput “omics” analysis with metabolite/protein/gene interaction structure, to sensory data from internet of things, to applications in astronomy and natural language processing. We investigate the role of structure by moving beyond the “graph” in graphical models, starting from discrete structures that appear in combinatorial problems from clustering and matching to packing, tiling and factorization. For data-intensive domains, most unsupervised (and semi-supervised) tasks such as anomaly detection and dimensionality reduction can be expressed as inference in a deep latent variable model, where domain structure can be expressed within the layered neural architecture.  After reviewing discrete examples in medical imaging and astronomy I will discuss recent progress in this area and future research directions.

Bio:    Siamak Ravanbakhsh is currently a Postdoctoral Research Fellow in the Robotics Institute and the Machine Learning Department at Carnegie Mellon University.  He is broadly interested in probabilistic reasoning in complex, structured and combinatorial domains and its applications in scientific discovery. He obtained his PhD from University of Alberta in 2015, as a member of Alberta Ingenuity Center for Machine Learning. He obtained his M.Sc. from the same university in 2009 and his B.Sc from Sharif University in Tehran.

For more information, contact:
Sandra Hall

Events are free and open to the public unless otherwise noted.