Sun 15 Jan 2023 12:20 - 12:25 at Scollay - Second Session Chair(s): Steven Holtzen, Christine Tasson

Probabilistic programming languages (PPLs) have emerged as a prominent area of research in recent years due to their ability to democratize probabilistic modeling. Current PPLs either do not support or do not scale well on real-life hybrid probabilistic programs. We present HyBit, an approximate inference algorithm with an aim to provide better support and scalability for hybrid programs. In HyBit, we first obtain a discrete abstraction of the probabilistic program by bitblasting the continuous distributions. Then we harness the power of existing discrete PPLs to perform exact inference on the discrete abstraction. This approach comes with the challenge of enumerating exponential number of values. To counter this problem, we present an efficient way to approximate a discrete distribution using linear piece-wise distributions which require enumerating values only linear in the number of pieces. In this work, we prove theoretically that as we increase the number of bits of the discrete abstraction, we get closer to the ground truth. We provide empirical evidence to show that bitblasting probabilistic programs is a practical approach to performing probabilistic inference.

Bit-Blasting Probabilistic Programs (lafi23-final47.pdf)285KiB

Sun 15 Jan

Displayed time zone: Eastern Time (US & Canada) change

11:00 - 12:30
Second SessionLAFI at Scollay
Chair(s): Steven Holtzen Northeastern University, Christine Tasson Sorbonne Université — LIP6
11:00
20m
Talk
What do posterior distributions of probabilistic programs look like?Boston
LAFI
Mathieu Huot University of Oxford, A: Alexander K. Lew Massachusetts Institute of Technology, Vikash K. Mansinghka Massachusetts Institute of Technology, Sam Staton University of Oxford
File Attached
11:20
10m
Talk
Semantics of Probabilistic Program TracesBoston
LAFI
Alexander K. Lew Massachusetts Institute of Technology, A: Eli Sennesh Northeastern University, Jan-Willem Van De Meent University of Amsterdam, Vikash Mansinghka Massachusetts Institute of Technology
File Attached
11:30
10m
Talk
A convenient category of tracing measure kernelsBoston
LAFI
A: Eli Sennesh Northeastern University, Jan-Willem Van De Meent University of Amsterdam
File Attached
11:45
5m
Talk
Random probability distributions as natural transformationsParis
LAFI
A: Victor Blanchi ENS Paris, Hugo Paquet University of Oxford
File Attached
11:50
5m
Talk
Static Delayed Sampling for Probabilistic Programming LanguagesParis
LAFI
A: Gizem Caylak KTH Royal Institute of Technology, Daniel Lundén KTH Royal Institute of Technology, Viktor Senderov Naturhistoriska riksmuseet, David Broman KTH Royal Institute of Technology
11:55
5m
Talk
Denotational semantics of languages for inference: semirings, monads, and tensorsOnline
LAFI
Cristina Matache University of Edinburgh, A: Sean K. Moss University of Cambridge, Sam Staton University of Oxford, Ariadne Si Suo University of Oxford
12:10
5m
Talk
Separated and Shared Effects in Higher-Order LanguagesBoston
LAFI
A: Pedro Henrique Azevedo de Amorim Cornell University, Justin Hsu Cornell University
12:15
5m
Talk
On Iteration in Discrete Probabilistic ProgrammingBoston
LAFI
A: Mateo Torres-Ruiz , Robin Piedeleu University of Oxford, Alexandra Silva Cornell University, Fabio Zanasi University College London
File Attached
12:20
5m
Talk
Bit-Blasting Probabilistic ProgramsBoston
LAFI
A: Poorva Garg University of California, Los Angeles, Steven Holtzen Northeastern University, Guy Van den Broeck University of California at Los Angeles, Todd Millstein University of California at Los Angeles
File Attached
12:25
5m
Talk
πMPC: Automatic Security Proofs for MPC ProtocolsBoston
LAFI
A: Mako P. Bates University of Vermont, Joseph P. Near University of Vermont