Probabilistic modeling and inference problems are often computationally intractable, so practitioners frequently design custom approximation algorithms for the problem at hand. However, current probabilistic programming languages typically provide only a small set of black-box inference algorithms that cannot be customized. This limitation renders most probabilistic programming languages unsuitable for most applications of probabilistic modeling and inference. This paper introduces novel inference meta-programming constructs for expressing custom inference algorithms in probabilistic programming languages. These constructs enable probabilistic programmers to (i) dynamically decompose inference problems into subproblems; (ii) apply inference tactics to subproblems; (iii) alternate between incorporating new data and performing inference given existing data; and (iv) explore multiple execution traces of the probabilistic program at once. Together, these inference meta-programming constructs allow probabilistic programmers to express a broad class of algorithms, including novel hybrids of gradient-based search, combinatorial search, Markov chain Monte Carlo, sequential Monte Carlo, and variational inference. Experimental results on a collection of probabilistic programs, written in the Venture probabilistic programming language, highlight the significant performance and accuracy benefits that programmable inference can deliver.
Fri 22 Jun Times are displayed in time zone: (GMT-04:00) Eastern Time (US & Canada) change
|11:00 - 11:25|
Marco Cusumano-TownerMIT-CSAIL, Benjamin BichselETH Zurich, Switzerland, Timon Gehr, Martin VechevETH Zürich, Vikash MansinghkaMITMedia Attached
|11:25 - 11:50|
Timon Gehr, Sasa MisailovicUniversity of Illinois at Urbana-Champaign, USA, Petar TsankovETH Zurich, Laurent VanbeverETH Zürich, Pascal WiesmannETH Zurich, Switzerland, Martin VechevETH ZürichMedia Attached
|11:50 - 12:15|
Vikash MansinghkaMIT, Ulrich SchaechtleMassachusetts Institute of Technology, USA, Shivam Handa, Alexey Radul, Yutian ChenGoogle Deepmind, n.n., Martin RinardMassachusetts Institute of TechnologyMedia Attached