CS Colloquium: Stephan Mandt (Disney Research) - Next generation variational inference: algorithms, models, and applications
Mon, Apr 03, 2017 @ 01:00 PM - 02:00 PM
Conferences, Lectures, & Seminars
Speaker: Stephan Mandt, Disney Research
Talk Title: Next generation variational inference: algorithms, models, and applications
Series: CS Colloquium
Abstract: This lecture satisfies requirements for CSCI 591: Computer Science Research Colloquium.
Probabilistic modeling is a powerful paradigm in machine learning. In this field, we assume a generative process in order to explain our observations, and then use a Bayesian inference algorithm to reason about its parameters. Probabilistic modeling has become scalable due to stochastic variational inference which reduces Bayesian inference to non-convex stochastic optimization. This talk focuses on two new inference algorithms: variational tempering-an algorithm that operates on several artificial temperatures simultaneously to find better local optima, and constant SGD-a scalable inference algorithm with applications to hyperparameter optimization. I will then present several new models that have become tractable due to modern variational inference with applications in text modeling, recommendations, and computer vision. I will show how a probabilistic view on Google's word2vec algorithm allows for extensions to other types of high dimensional data and show new applications: analyzing supermarket shopping data, movie ratings, and tracking semantic changes of individual words over centuries of digitized books. Finally, I will show how factorized variational autoencoders allow us to analyze audience reactions to movies.
Biography: Stephan Mandt is a research scientist at Disney Research Pittsburgh, where he leads the statistical machine learning group. From 2014 to 2016 he was a postdoctoral researcher with David Blei at Columbia University, and from 2012 to 2014 a PCCM postdoctoral fellow at Princeton University. Stephan did his Ph.D. with Achim Rosch at the Institute for Theoretical Physics at the University of Cologne, supported by a fellowship of the German National Merit Foundation. His research interests include scalable approximate Bayesian inference and machine learning for media analytics.
Host: Fei Sha
Location: Kaprielian Hall (KAP) - 140
Audiences: Everyone Is Invited
Contact: Assistant to CS chair