Structure Learning for Sum-Product Networks

KEC 1003
Mon, 10/19/2015 - 4:00pm

Daniel Lowd
Assistant Professor, Department of Computer and Information Science,
University of Oregon

Abstract:
Probabilistic graphical models have been applied to many domains, including
computer vision, natural language processing, and bioinformatics. However,
their effectiveness is limited by the complexity of inference, which is
generally intractable. An appealing alternative is to work with tractable
probabilistic models, in which exact inference is efficient. Sum-product
networks (SPNs) are a deep, tractable probabilistic representation that
generalize many other tractable model classes. SPNs have achieved
state-of-the-art results on computer vision and density estimation problems,
but selecting a good structure for an SPN is challenging.

In this talk, I will provide a brief introduction to SPNs and then discuss
several recent approaches to learning SPN structures from data. The first
approach is to adapt standard graphical model structure learning algorithms,
resulting in SPNs that represent tractable graphical models. The second
approach is to recursively cluster instances and variables, resulting in SPNs
that represent hierarchical mixture models. These two approaches capture
different types of patterns in the data. The ID-SPN algorithm uses a combined
approach, leading to much better structures on a variety of benchmark
domains. In many cases, ID-SPN learns SPNs that are more accurate than
intractable Bayesian networks, demonstrating that SPNs can maintain
tractability without sacrificing accuracy.

Bio:


URL:
http://eecs.oregonstate.edu/colloquium/structure-learning-sum-product-networks

_______________________________________________
Colloquium mailing list
[email protected]
https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium

Reply via email to