SciPost Submission Page
New Machine Learning Techniques for Simulation-Based Inference: InferoStatic Nets, Kernel Score Estimation, and Kernel Likelihood Ratio Estimation
by Kyoungchul Kong, Konstantin T. Matchev, Stephen Mrenna, Prasanth Shyamsundar
This is not the latest submitted version.
This Submission thread is now published as
Submission summary
| Authors (as registered SciPost users): | Kyoungchul Kong · Konstantin Matchev · Stephen Mrenna · Prasanth Shyamsundar |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2210.01680v1 (pdf) |
| Code repository: | https://gitlab.com/prasanthcakewalk/code-and-data-availability/ |
| Data repository: | https://gitlab.com/prasanthcakewalk/code-and-data-availability/ |
| Date submitted: | 2022-10-11 22:59 |
| Submitted by: | Shyamsundar, Prasanth |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Computational, Phenomenological |
Abstract
We propose an intuitive, machine-learning approach to multiparameter inference, dubbed the InferoStatic Networks (ISN) method, to model the score and likelihood ratio estimators in cases when the probability density can be sampled but not computed directly. The ISN uses a backend neural network that models a scalar function called the inferostatic potential $\varphi$. In addition, we introduce new strategies, respectively called Kernel Score Estimation (KSE) and Kernel Likelihood Ratio Estimation (KLRE), to learn the score and the likelihood ratio functions from simulated data. We illustrate the new techniques with some toy examples and compare to existing approaches in the literature. We mention en passant some new loss functions that optimally incorporate latent information from simulations into the training procedure.
Current status:
Reports on this Submission
Report #1 by Ramon Winterhalder (Referee 1) on 2022-11-8 (Invited Report)
- Cite as: Ramon Winterhalder, Report on arXiv:2210.01680v1, delivered 2022-11-08, doi: 10.21468/SciPost.Report.6085
Strengths
The usage of a "backend" function phi, which obeys several useful properties as shown in eq (10) instead of learning the score and the ratio directly, shows great potential. Furthermore, baking-in symmetries into the network architecture that are fulfilled precisely and do not have to be learned are a great idea, as it generally increases the precision and, most importantly, the stability of the network.
Weaknesses
Missing explicit LHC examples.
Report
Thanks to the author for the interesting paper. The presented approaches and the results are exciting and offer great potential for future applications for LHC analyses. For this reason, the article is worth publishing. However, as the paper does not explicitly showcase any physics example, I would either ask the others to consider some LHC examples and add them to their draft or recommend the paper for the partner journal SciPost Physics Codebases. Further, I would ask the authors to add another citation (see the requested changes).
Requested changes
Before publishing, I would ask the others to cite another method of directly estimating the likelihood using the Matrix-Element Method (2210.00019), which would fit into Table 1 next to the NDE approach.
Author: Prasanth Shyamsundar on 2023-02-14 [id 3353]
(in reply to Report 1 by Ramon Winterhalder on 2022-11-08)Hey Ramon,
Thanks a lot for your feedback and suggestions! We have added the citation as suggested.
Several papers in the last few years have already showcased the applications of score and likelihood ratio estimation in particle physics. We believe that redoing physics examples will only lead to duplication of effort, without significantly improving the value of the manuscript, which is already very lengthy as it is. We'll leave the decision of the appropriate journal to the editor.
Regards,
Prasanth (on behalf of the authors)