Fri 12 Jan 2018 13:40 - 14:05 at Watercourt - Probability Chair(s): Lars Birkedal

Program sensitivity, also known as Lipschitz continuity, describes how small changes in a program’s input lead to bounded changes in the output. We propose an average notion of program sensitivity for probabilistic programs—expected sensitivity—that averages a distance function over a probabilistic coupling of the two output distributions from two nearby inputs. By varying the distance functions, expected sensitivity captures useful notions of probabilistic function sensitivity, including algorithmic stability of machine learning algorithms and convergence of Markov chains.

Furthermore, expected sensitivity satisfies clean compositional properties and is amenable to formal verification. We develop a relational program logic called ExSeL for proving expected sensitivity properties, featuring two key components. First, relational pre-conditions and post-conditions are expressed using distances, which can be seen as a real-valued generalization of typical boolean-valued (relational) assertions. Second, judgments —with distances as pre-conditions and post-conditions— are interpreted in terms of expectation coupling, a novel, quantitative generalization of probabilistic couplings which supports compositional reasoning.

We demonstrate our logic on two classes of examples: uniform stability of the Stochastic Gradient Method algorithm from machine learning, and rapid mixing for a model of asexual population dynamics. We also extend our logic with a transitivity principle for expectation couplings to capture the path coupling, and we demonstrate our extension by proving rapid mixing of the Glauber dynamics from statistical physics.

Fri 12 Jan

POPL-2018-papers
13:40 - 15:20: Research Papers - Probability at Watercourt
Chair(s): Lars BirkedalAarhus University
POPL-2018-papers151576080000013:40 - 14:05
Talk
Gilles BartheIMDEA Software Institute, Thomas EspitauUniversite Pierre et Marie Curie, Benjamin GregoireINRIA, Justin HsuUniversity College London, Pierre-Yves StrubEcole Polytechnique
POPL-2018-papers151576230000014:05 - 14:30
Talk
Aws AlbarghouthiUniversity of Wisconsin-Madison, Justin HsuUniversity College London
POPL-2018-papers151576380000014:30 - 14:55
Talk
Thomas EhrhardCNRS and University Paris Diderot, Michele PaganiUniversity Paris Diderot, Christine TassonUniversity Paris Diderot
POPL-2018-papers151576530000014:55 - 15:20
Talk
Adam ŚcibiorUniversity of Cambridge and MPI Tuebingen, Ohad KammarUniversity of Oxford, Matthijs VákárUniversity of Oxford, Sam StatonUniversity of Oxford, Hongseok YangUniversity of Oxford, Yufei CaiUniversity of Tuebingen, Klaus OstermannUniversity of Tuebingen, Sean K. MossUniversity of Cambridge, Chris HeunenUniversity of Edinburgh, Zoubin GhahramaniUniversity of Cambridge