|
AAAI 2018 Tutorial Integrating Learning Into Reasoning
The tutorial is aimed at
researchers in KRR who are interested in how to best use machine
learning to generate knowledge bases that contain knowledge that
is hard to express by hand. Machine learning has been very
successful at some tasks of this kind, and in many cases there is
currently no alternative. The tentative nature of learned knowledge
requires a semantics that is less strict than classical logics, and
this tutorial will introduce participants to PAC-Semantics, a
particularly simple suitable choice. In particular, in this formal
framework, there are quantifiable benefits to integrating the learning
and reasoning processes. We will explain what these benefits are, and
illustrate some simple techniques for such integrated learning and
reasoning and their analysis. Participants should come away from the
tutorial with the ability to establish that an algorithm’s output is
sound in a PAC sense, using standard arguments based on bounds on
representation size and “concentration of measure” from independent
examples. In addition, PAC-Semantics and
these integrated learning and reasoning methods offer some new
perspectives on classical issues in KRR, such as nonmonotonic
reasoning, the qualification problem, and elaboration tolerance. The
framework also naturally yields a new formulation of abductive
inference. We will explain these connections, but also highlight the
issues that we believe are not addressed and still have substantial
scope for interesting KRR work. Participants should come away with an
appreciation for how features such as error tolerance in learning and
conditional probability together give rise to nonmonotonic effects
under the PAC-Semantics, as well as some interesting open questions in
the area that they might be well positioned to address. Download
the slides for the tutorial (mirror 1).
Presenters' Short Bio: Brendan Juba is an Assistant Professor in the Department of Computer Science and Engineering at Washington University in St. Louis. His current research interests lie in theoretical approaches to Artificial Intelligence, founded on the theory of Algorithms and Computational Complexity. He is also interested in Theoretical Computer Science more broadly construed. Previously, Brendan worked as a postdoc under the supervision of Leslie Valiant, jointly affiliated with Harvard and MIT with the Center for Science of Information until the fall of 2012, and subsequently solely affiliated with Harvard through summer 2014. He completed his Ph.D. at MIT in 2010 under the supervision of Madhu Sudan, and his dissertation, "Universal Semantic Communication" was published by Springer in 2011. Brendan also holds a M.S. in Mathematical Sciences and B.S. in Computer Science from Carnegie Mellon University, both awarded in 2005. His work is currently supported by a 2015 AFOSR Young Investigator Award. Loizos Michael is an Assistant
Professor at Open University of Cyprus, where he founded and directs
the Computational Cognition Lab, and also serves as the academic head of the
cross-institutional M.Sc. Program in Cognitive Systems
(http://cogsys.ouc.ac.cy). His research focuses on the development of
formal computational models for cognitive processes associated with
individual or collective intelligence, using techniques from
computational learning theory, logic-based knowledge representation,
and commonsense reasoning. Loizos received a B.Sc. in Computer Science with a minor degree in Mathematics from University of Cyprus, graduating
top of the class of 2002, and receiving the Republic of Cyprus
Presidential Award. He continued his education at Harvard University,
where he received an M.Sc. and a Ph.D. in Computer Science in 2008,
under the supervision of Leslie Valiant. Before joining Open University
of Cyprus in 2009, he held a visiting faculty appointment at University
of Cyprus. He served as the PC chair of the 15th European Conference on
Logics in Artificial Intelligence (JELIA 2016), and he has been
organizing the workshop series on Cognitive Knowledge Acquisition and
Applications (http://cognitum.ws) since 2015. |