SciPost Submission Page
Resonance Searches with Machine Learned Likelihood Ratios
by Jacob Hollingsworth and Daniel Whiteson
Submission summary
| Authors (as registered SciPost users): | Jacob Hollingsworth |
| Submission information | |
|---|---|
| Preprint Link: | scipost_202003_00050v3 (pdf) |
| Date submitted: | 2021-08-11 16:25 |
| Submitted by: | Hollingsworth, Jacob |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approach: | Phenomenological |
Abstract
We demonstrate the power of machine-learned likelihood ratios for resonance searches in a benchmark model featuring a heavy Z' boson. The likelihood ratio is expressed as a function of multivariate detector level observables, but rather than being calculated explicitly as in matrix-element-based approaches, it is learned from a joint likelihood ratio which depends on latent information from simulated samples. We show that bounds drawn using the machine learned likelihood ratio are tighter than those drawn using a likelihood ratio calculated from histograms.
Author comments upon resubmission
With regards to the question of assigning uncertainties to calculated p-values, it is our understanding that this is a current topic of research surrounding these methods. Ideas for achieving this include training an ensemble of machine learning models to predict the likelihood ratio, rather than a single neural network. The ensemble mean (or alternative method of ensemble prediction) would then be the predicted likelihood ratio, and the ensemble variance could be used to quantify uncertainty in the likelihood ratio, and thus the p-values, due to the probabilistic training of ensemble members.
List of changes
We have included a new paragraph beginning "Built into these approaches..." on page 4.
Current status:
Reports on this Submission
Report
Thank you to the authors for taking into account my feedback! I am happy to recommend publication. The only last comment I have is a quick followup to the response:
"With regards to the question of assigning uncertainties to calculated p-values...Ideas for achieving this include training an ensemble ... the ensemble variance could be used to quantify uncertainty in the likelihood ratio..."
This would give you a sense for a statistical component of the uncertainty, but it does not tell you about the potential bias from the training procedure or network architecture. Perhaps it is worth adding a sentence about this (I agree it is an active area of research).