BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Moritz Schauer (Chalmers University of Technology & University of Gothenburg) DTSTART:20221020T131500Z DTEND:20221020T140000Z DTSTAMP:20260422T155025Z UID:gbgstats/5 DESCRIPTION:Title: Automatic differentiation of programs with discrete randomness\nby Mo ritz Schauer (Chalmers University of Technology & University of Gothenburg ) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\n\nAb stract\nAutomatic differentiation (AD)\, a technique for constructing new programs which compute the derivative of an original program\, has become ubiquitous throughout scientific computing and deep learning due to the im proved performance afforded by gradient-based optimization. However\, AD s ystems have been restricted to the subset of programs that have a continuo us dependence on parameters. Programs that have discrete stochastic behavi ors governed by distribution parameters\, such as flipping a coin with pro bability p of being heads\, pose a challenge to these systems because the connection between the result (heads vs tails) and the parameters (p) is f undamentally discrete. In this paper we develop a new reparameterization-b ased methodology that allows for generating programs whose expectation is the derivative of the expectation of the original program. We showcase how this method gives an unbiased and low-variance estimator which is as auto mated as traditional AD mechanisms. We demonstrate unbiased forward-mode A D of discrete-time Markov chains\, agent-based models such as Conway's Gam e of Life\, and unbiased reverse-mode AD of a particle filter. Our code is available at https://github.com/gaurav-arya/StochasticAD.jl\n LOCATION:https://researchseminars.org/talk/gbgstats/5/ END:VEVENT END:VCALENDAR