- A Pseudo-Likelihood Function for Relational Data. A
fundamental difficulty in learning Bayes Nets for relational or network
data is to define a likelihood function that measures how well a
statistical model fits a database. The reason this is difficult is that
the usual product likelihood function assumes that population units are
independent, whereas in relational data, by definition units, are
interdependent. My proposed solution is to consider the expected
likelihood of a model for small, randomly selected subgroups of units.
Here are two talks presenting this idea, for different audiences.
- Discovery of Conservation Laws.
Oct 2010. PDF Version.
Intended Audience: Computer Scientists. Given at Discovery Science
2010 conference. Presents a new criterion for selecting conservation
laws: they should be maximally simple and maximally strict. Formally,
minimize the L1-norm subject to the constraint that they rule out as
many unobserved reaction as possible. We provide an efficient
optimization algorithm for this new criterion. The procedure rediscovers
conservation principles in molecular chemistry and in the fundamental
Standard Model of Particle Physics.
- Causal Modelling for Relational Data.
July 2010. PDF Version.
Intended Audience: Logicians, Computer Scientists, Epistemologists.
Given at CMU. An outline of what I see as the main difficulties in
statistical-relational learning. Presents briefly our new learn-and-join
algorithm for Bayes net structure learning with relational data. Also a
new semantics for relational Bayes nets, based on random groundings
rather than a ground model.
- The IMAP Hybrid Method for Learning Gaussian Bayes Nets.
May 2010. PDF Version.
Intended Audience: Computer Scientists, Statisticians. A conference
talk on hybrid methods for learning Bayes nets. Presents
constraint-based methods, score-based methods, and a new way to combine
them. This paper won the best paper award - thank you CanAI!
- Three Applications of Means-Ends Epistemology.
Intended Audience: Philosophers and Computer Scientists (given at
Caltech Philosophy Seminar). Presents three applications of formal
learning theory as a package: Discovering conservation laws in particle
physics, solving Goodman's Riddle of Induction and learning causal
models from observed correlations.
- Mind Change Optimal Learning of Bayes Net Structure.
Intended Audience: Computer Scientists. Presents our results on the
mind change complexity of learning Bayes nets from correlation data.
Contains a simple example of the mind change optimal learning method
(minimizing the number of edges), and a brief outline of the NP-hardness
proof for this method.
- Evolutionary Equilibria in Network Routing Games: Specialization and Clustering. Nov 2005.
Intended Audience: Computer Scientists. Joint work with Petra
Berenbrink. Presents a model of network traffic where users choose the
routes for their messages by themselves. We analyze the evolutionary
equilibria of such games, and find that evolution leads to the formation
of "niches" in which links "specialize" in certain tasks. These network
games are closely related to congestion games.
- The Evidence for Conservation Laws and Particle Families. May 2005.
Intended Audience: Philosophers of Physics, Physicists. Similar to
previous talk, but with more technical details. Also a brief overview of
the methods for inferring the existence of unobserved particles and the
application to the question of whether the electron neutrino is a
Majorana particle. University of Western Ontario.
- How Particle Physics Cuts Nature At Its Joints. May 2005.
Intended Audience: Philosophers of Science. Discusses the results of
my algorithms and analysis of finding conservation laws corresponding
to particle families. With reference to traditional philosophical issues
such as skepticism, underdetermination, nominalism, natural kinds, laws
of nature. University of Washington.
- Learning Conservation Principles in Particle Physics. March 2005.
Intended Audience: Computer Scientists. A brief overview of my
algorithms for finding conserved quantities and hidden particles in
particle reactions. Laboratory for Computational Intelligence,
University of British Columbia.
- Automated Search for Conserved Quantities in Particle Reactions. February 2005.
Audience: Physicists. An overview of most of my results on algorithms
for finding conserved quantities and hidden particles in particle
reactions - see NSERC project. Particle Physics Group, Simon Fraser