Max Welling
Wednesday 15th April 2015
Time: 4.00pm
Basement Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Bayesian Inference in Complex Generative Models
In a time when deep learning and big data claim the center
stage in machine learning, generative models and Bayesian inference have
moved somewhat out of the limelight. I argue that generative modeling
and Bayesian learning will remain key for many exciting applications. I
present three advances in this direction developed in my group: 1) a new
variational learning algorithm for the Helmholtz machine (the
variational auto-encoder) applied to semi-supervised learning, 2) a new
large-scale distributed posterior MCMC sampling procedure applied to
Matrix Factorization and 3) an efficient posterior MCMC sampling
algorithms for complex, likelihood free simulator models. I will
conclude with a reflection on what is still missing to build truly large
scale, distributed and privacy preserving learning systems.
Joint work with Ted Meeds, Durk Kingma and Sungjin Ahn.