We present new algorithms for unsupervised learning of probabilistic topic models and noisy-OR Bayesian networks. Probabilistic topic models are frequently used to learn thematic structure from large document collections without human supervision, and the Bayesian networks that we study are often used for medical diagnosis. We circumvent the computational intractability of maximum likelihood learning by making the assumption that the observed data is drawn from a distribution within the model family that we are attempting to learn, such as Bayesian networks with latent variables. We demonstrate a set of structural constraints that make learning possible, yet are still realistic for many real-world applications. The new algorithms produce results comparable to the best MCMC implementations while running orders of magnitude faster.
(Joint work with Sanjeev Arora, Rong Ge, Yoni Halpern, David Mimno, Ankur Moitra, Yichen Wu, and Michael Zhu)