Calibrating Neural Simulation-Based Inference with Differentiable Coverage Probability

Saved in:
Bibliographic Details
Title: Calibrating Neural Simulation-Based Inference with Differentiable Coverage Probability
Authors: Falkiewicz, Maciej, Takeishi, Naoya, Shekhzadeh, Imahn, Wehenkel, Antoine, Delaunoy, Arnaud, Louppe, Gilles, Kalousis, Alexandros
Source: Advances in Neural Information Processing Systems (2023-12); Advances in Neural Information Processing Systems, New Orleans, United States - Louisiana [US-LA], December 2023
Publication Year: 2023
Subject Terms: Statistics - Machine Learning, Computer Science - Learning, Engineering, computing & technology, Computer science, Ingénierie, informatique & technologie, Sciences informatiques
Description: Bayesian inference allows expressing the uncertainty of posterior beliefunder a probabilistic model given prior information and the likelihood of theevidence. Predominantly, the likelihood function is only implicitly establishedby a simulator posing the need for simulation-based inference (SBI). However,the existing algorithms can yield overconfident posteriors (Hermans *et al.*,2022) defeating the whole purpose of credibility if the uncertaintyquantification is inaccurate. We propose to include a calibration term directlyinto the training objective of the neural model in selected amortized SBItechniques. By introducing a relaxation of the classical formulation ofcalibration error we enable end-to-end backpropagation. The proposed method isnot tied to any particular neural model and brings moderate computationaloverhead compared to the profits it introduces. It is directly applicable toexisting computational pipelines allowing reliable black-box posteriorinference. We empirically show on six benchmark problems that the proposedmethod achieves competitive or better results in terms of coverage and expectedposterior density than the previously existing approaches.
Document Type: conference paper
http://purl.org/coar/resource_type/c_5794
conferenceObject
peer reviewed
Language: English
Relation: urn:issn:1049-5258
DOI: 10.48550/arXiv.2310.13402
Access URL: https://orbi.uliege.be/handle/2268/308498
Rights: open access
http://purl.org/coar/access_right/c_abf2
info:eu-repo/semantics/openAccess
Accession Number: edsorb.308498
Database: ORBi
Description
Abstract:Bayesian inference allows expressing the uncertainty of posterior beliefunder a probabilistic model given prior information and the likelihood of theevidence. Predominantly, the likelihood function is only implicitly establishedby a simulator posing the need for simulation-based inference (SBI). However,the existing algorithms can yield overconfident posteriors (Hermans *et al.*,2022) defeating the whole purpose of credibility if the uncertaintyquantification is inaccurate. We propose to include a calibration term directlyinto the training objective of the neural model in selected amortized SBItechniques. By introducing a relaxation of the classical formulation ofcalibration error we enable end-to-end backpropagation. The proposed method isnot tied to any particular neural model and brings moderate computationaloverhead compared to the profits it introduces. It is directly applicable toexisting computational pipelines allowing reliable black-box posteriorinference. We empirically show on six benchmark problems that the proposedmethod achieves competitive or better results in terms of coverage and expectedposterior density than the previously existing approaches.
DOI:10.48550/arXiv.2310.13402