[Pl-seminar] Daniel Huang -- Augur: Data-Parallel Probabilistic Modeling. Feb. 20th, 11:45am

[William J. Bowman]

Hello all,

We will have a our first PL seminar on Feb. 20th, 11:45am in 366 WVH:

Speaker: Daniel Huang

Title: Augur: Data-Parallel Probabilistic Modeling

Abstract: Probabilistic programming systems automatically generate an
inference algorithm from a description of a model as a program. The goal is
to make it easier for practitioners to apply Bayesian machine learning.
However, a poorly performing inference algorithm can make this approach
impractical. In this talk, we describe Augur, a system that extracts a
data-parallel, single-chain, MCMC inference algorithm from a restricted
class of models for execution on a GPU. The focus of the project is on
performance of the inference algorithm, instead of expressiveness of the
modeling language. We detail our approach and highlight some of the
challenges we faced in generating performant code.


-- 
William J. Bowman
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://lists.ccs.neu.edu/pipermail/pl-seminar/attachments/20150215/55071085/attachment.pgp>