BEGIN:VCALENDAR VERSION:2.0 PRODID:researchseminars.org CALSCALE:GREGORIAN X-WR-CALNAME:researchseminars.org BEGIN:VEVENT SUMMARY:Umberto Picchini (Chalmers University of Technology & University o f Gothenburg) DTSTART:20240515T111500Z DTEND:20240515T120000Z DTSTAMP:20260422T155052Z UID:gbgstats/54 DESCRIPTION:Title: Fast\, lightweight and semi-amortised simulation-based inference\nby Umberto Picchini (Chalmers University of Technology & University of Gothe nburg) as part of Gothenburg statistics seminar\n\nLecture held in MVL14.\ n\nAbstract\nBayesian inference for complex models with an intractable lik elihood can be tackled using algorithms performing many calls to computer simulators. These approaches are collectively known as "simulation-based i nference" (SBI). Recent SBI methods use neural-conditional-estimation\, th at is neural networks are employed to provide approximations to the likeli hood function or the posterior distribution of model parameters. While neu ral-based posterior and likelihood estimation methods have produced except ionally flexible inference strategies\, these can be computationally inten sive to run and have a non-negligible impact on energy expenditure and mem ory requirements. In this work\, rather than using neural networks we prop ose more "frugal" strategies that display state-of-art inference quality\, while being able to run with limited resources\, being much faster to tra in and exhibiting a much smaller computational footprint. We investigate s tructured mixtures of probability distributions and design a new SBI metho d named Sequential Mixture Posterior and Likelihood Estimation (SeMPLE). S eMPLE learns closed-form approximations for both the posterior $p(θ|y)$ a nd the likelihood $p(y|θ)$ from the same training data\, using Gaussian m ixture models that can be efficiently learned.\nWe show favorable results for a variety of stochastic models (including SDEs and Markov jump process es)\, also in presence of multimodal posteriors. \n\nThe talk will be appr oachable for the uninitiated audience\, while novel results will be of int erest for the experienced audience.\n\nJoint work with Henrik Häggström\ , Pedro L. C. Rodrigues\, Geoffroy Oudoumanessah and Florence Forbes\, htt ps://arxiv.org/abs/2403.07454\n LOCATION:https://researchseminars.org/talk/gbgstats/54/ END:VEVENT END:VCALENDAR