Pedro Domingos, University of Washington
Large-Scale Learning and Inference: What We Have Learned with Markov Logic Networks
Markov logic allows very large and rich graphical
models to be compactly specified. Current learning and inference algorithms for
Markov logic can routinely handle models with millions of variables, billions of features,
thousands of latent variables, and strong dependencies. In this talk I
will give an overview of the main ideas in these algorithms, including
weighted satisfiability, MCMC with deterministic dependencies, lazy
inference, lifted inference, relational cutting planes, scaled conjugate
gradient, relational clustering and relational pathfinding. I will also
discuss the lessons learned in developing successive generations of these
algorithms and promising ideas for the next round of scaling up.
(Joint work with Stanley Kok, Daniel Lowd, Hoifung Poon, Matt Richardson, Parag Singla, Marc Sumner, and Jue Wang.)