Abstract:
Collective graphical models (CGMs) are a formalism for inference and learning
with (noisy) aggregate data that are motivated by a model for bird migration. We
discuss the problem of approximate MAP inference in CGMs. We first describe
how MAP inference can be formulated as a convex optimization problem using
two approximations, and then derive a novel Belief Propagation (BP)-style
algorithm for this convex optimization problem. The algorithm is a strict
generalization of BP. We show using synthetic datasets that this combination of
techniques speeds up inference and learning in CGMs by at least two orders
of magnitude compared with the previous state-of-the-art while providing
solutions of equal or better quality. We then show how this inference algorithm
can be used in a CGM model for bird migration to reconstruct migration routes
from birdwatcher reports.
Bio:
Tao Sun is a PhD student in the School of Computer science, UMass
Amherst, advised by Dan Sheldon. He is a member of Machine Learning
and Data Science (MLDS) Lab.
He is interested in machine learning and its applications, and is now
working on the project BirdCast to develop novel methods for
continent-scale bird migration prediction.