One World ABC Seminar: Pierre-Alexandre Mattei and Samuel Wiqvist

on the September 3, 2020

at 11:30 am [UK time]
For this ninth session of the One World ABC Seminar, Pierre-Alexandre Mattei from Inria and Samuel Wiqvist from Lund university will talk about "Partially Exchangeable Networks and Architectures for Learning Summary Statistics in Approximate Bayesian Computation".

Inspired by the "One World Probability Seminar", we decided to run The One World ABC Seminar, a weekly/fortnightly series of seminars that will take place on Blackboard Collaborate on Thursdays at 11.30am [UK time]. The idea is to gather members and disseminate results and innovation during these weeks and months under lockdown.

Pierre-Alexandre Mattei and Samuel Wiqvist

In this talk, we will introduce partially exchangeable networks (PENs) [1], and we will, inparticular, discuss the application of learning summary statistics for approximate Bayesiancomputation (ABC). Connections between our methodology and other deep learning-basedmethods  for  simulation  inference  will  also  be  highlighted.   By  design,  PENs  are  invariantto  block-switch  transformations,  which  characterize  the  partial  exchangeability  propertiesof conditionally Markovian processes.  More- over, we show that any block-switch invariantfunction has a PEN-like representation.  The DeepSets architecture is a special case of PENand we can therefore also target fully exchangeable data. We employ PENs to learn summarystatistics in ABC. When comparing PENs to previous deep learning methods for learningsummary  statistics,  our  results  are  highly  competitive,  both  considering  time  series  andstatic models.  In- deed, PENs provide more reliable posterior samples even when using lesstraining data.This is joint work with Umberto Picchini and Jes Frellsen.
[1]  S. Wiqvist, P.-A. Mattei, U. Picchini, and J. Frellsen. Partially exchangeable networksand architectures for learning summary statistics in approximate Bayesian computation.In Proceedings of the 36th International Conference on Machine Learning. PMLR, 2019.
Published on September 1, 2020