[OxCSML Seminar] Sifan Liu (Stanford), Friday June 23 at 2pm, ​An Exact Sampler for Inference after Polyhedral Model Selection

[Jun Yang]

Hello everyone,

This Friday we will have our last OxCSML seminar in Trinity term.

Please see below the detailed information. Looking forward to seeing you on Friday at 2 pm!

Best,

Jun

Speaker: Sifan Liu (Stanford University)

Time and date: 23 June, 14.00 - 15.30

Place: Room LG.03, Department of Statistics, University of Oxford.

Title: An Exact Sampler for Inference after Polyhedral Model Selection

Abstract: Inference after model selection poses computational challenges when faced with intractable conditional distributions. Markov chain Monte Carlo (MCMC) is a common method for sampling from these distributions, but its slow convergence often limits its practicality. In this work, we propose a Monte Carlo sampler specifically designed for selective inference where the selection event can be characterized by a polyhedron. The method is based on importance sampling from a carefully chosen proposal distribution. Further variance reduction is achieved by conditional Monte Carlo and randomized quasi-Monte Carlo. Compared to MCMC, the proposed p-value estimator is unbiased, highly-accurate, and equipped with an error bound. Moreover, we present an approach to test and construct confidence intervals for multiple parameters using only a single batch of samples, reducing the need for repeated sampling. Numerical results demonstrate the efficiency of the proposed method compared to other alternatives for selective inference.

Zoom registration link: 

https://www.eventbrite.co.uk/e/oxcsml-seminar-23-june-2023-sifan-liu-stanford-university-tickets-658276441557