• View
• Edit
• History
• Print

Optimizing Trade-offs In Learning Graphical Models

Abstract: I will discuss trade-offs in learning probabilistic graphical models. I will begin by presenting sample complexity analysis for learning the parameters of Markov and Conditional Random Fields (MRFs and CRFs). Using composite likelihood, which generalizes likelihood and pseudolikelihood, one can derive PAC-style bounds. Analyzing these bounds reveals a full range of trade-offs between sample complexity, computational complexity, and potential for parallelism. I will demonstrate the possibility of tailoring composite likelihood estimators to problems in order to optimize these trade-offs. This talk is partly based on work presented at AISTATS 2012.

Speaker Bio: Joseph Bradley is a Ph.D. candidate at Carnegie Mellon University, advised by Carlos Guestrin. He works on learning large-scale Conditional Random Fields, particularly on methods which decompose problems to take advantage of parallel computation. Previously, he received a B.S.E. in Computer Science from Princeton University.