Stephan Mandt

(University of California)
hosted by Machine Learning Group of Prof. Marius Kloft

"Variational Inference and Probabilistic Embeddings"

Representation learning based on neural embeddings has experienced asurge in popularity. In this talk, I analyze word embeddings, knowledgegraph embeddings, and image embeddings through the lens of probabilisticmodeling. I will first summarize my group’s activities in creating andexploring probabilistic formulations and extensions of these models,such as dynamic word embeddings for language evolution, augmentedknowledge graph embeddings for link prediction, as well as sequentialvariational autoencoders for video prediction and compression. In thesecond part, I will talk about improving variational inference as apopular training paradigm for neural probabilistic modeling, where Ifocus on perturbative variational bounds, iterative amortized inference,and Quasi-Monte Carlo variational inference.

Bio: Stephan Mandt is an Assistant Professor at the University of California,Irvine. From 2016 to 2018, he led the statistical machine learninggroups at Disney Research Pittsburgh and Los Angeles. Stephan was apostdoctoral researcher with David Blei at Columbia and PrincetonUniversity. He obtained a Ph.D. in theoretical physics in 2012 from theUniversity of Cologne, supported by the German National MeritFoundation. He serves as an Area Chair NeurIPS, ICML, and AAAI.Stephan's interests include deep probabilistic modeling, variationalinference, and applications in the sciences and digital media.


Time: Wednesday, 14.08.2019, 16:15
Place: 42/110
Video: