Data Analysis in High-Energy Physics as a Differentiable Program
(2023)- Abstract
- Machine learning methods are now ubiquitous in physics, but often target objectives that are one or two steps removed from our physics goals. A prominent example of this is the discrimination between signal and background processes, which doesn’t account for the presence of systematic uncertainties – something crucial for the calculation of quantities such as the discovery significance and upper limits.
To combat this, this thesis shows that physics analysis workflows can be optimized in an end-to-end fashion, including the treatment of nuisance parameters that model systematic uncertainties, provided that the workflow is differentiable. By leveraging automatic differentiation and surrogates for non-differentiable operations, this... (More) - Machine learning methods are now ubiquitous in physics, but often target objectives that are one or two steps removed from our physics goals. A prominent example of this is the discrimination between signal and background processes, which doesn’t account for the presence of systematic uncertainties – something crucial for the calculation of quantities such as the discovery significance and upper limits.
To combat this, this thesis shows that physics analysis workflows can be optimized in an end-to-end fashion, including the treatment of nuisance parameters that model systematic uncertainties, provided that the workflow is differentiable. By leveraging automatic differentiation and surrogates for non-differentiable operations, this work has made this possible for the first time, and demonstrates its use in a proof-of-concept scenario.
This thesis will motivate the use of end-to-end optimization as described above, cover the techniques that make it possible, and show recent developments in a high-energy physics context. Future directions that aim to scale and apply these methods will also be highlighted.
In addition to this, a method to interpolate between the signatures of new physics models is presented, which uses normalizing flows. The thesis then goes on to show the use of the technique in a search for a new scalar boson 𝑆 produced in association with a Higgs boson from a heavy new scalar 𝑋. There are also some contributions that interpolate between the event yields with Gaussian processes, and that show how we can use normalizing flows to construct a likelihood ratio-inspired observable. (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/ea6b4b24-d2ea-41ae-8381-4ca4013d548a
- author
- Simpson, Nathan LU
- supervisor
-
- Caterina Doglioni LU
- Else Lytken LU
- opponent
-
- Professor Lista, Luca, University of Naples Federico II
- organization
- publishing date
- 2023
- type
- Thesis
- publication status
- published
- subject
- keywords
- differentiable programming, machine learning, particle physics
- pages
- 234 pages
- publisher
- Lund University
- defense location
- Rydbergsalen, Fysicum.
- defense date
- 2023-02-03 13:15:00
- ISBN
- 978-91-8039-494-9
- 978-91-8039-493-2
- DOI
- 10.5281/zenodo.7520315
- language
- English
- LU publication?
- yes
- id
- ea6b4b24-d2ea-41ae-8381-4ca4013d548a
- date added to LUP
- 2023-01-10 10:09:55
- date last changed
- 2023-01-11 14:54:55
@phdthesis{ea6b4b24-d2ea-41ae-8381-4ca4013d548a, abstract = {{Machine learning methods are now ubiquitous in physics, but often target objectives that are one or two steps removed from our physics goals. A prominent example of this is the discrimination between signal and background processes, which doesn’t account for the presence of systematic uncertainties – something crucial for the calculation of quantities such as the discovery significance and upper limits.<br/><br/>To combat this, this thesis shows that physics analysis workflows can be optimized in an end-to-end fashion, including the treatment of nuisance parameters that model systematic uncertainties, provided that the workflow is differentiable. By leveraging automatic differentiation and surrogates for non-differentiable operations, this work has made this possible for the first time, and demonstrates its use in a proof-of-concept scenario.<br/><br/>This thesis will motivate the use of end-to-end optimization as described above, cover the techniques that make it possible, and show recent developments in a high-energy physics context. Future directions that aim to scale and apply these methods will also be highlighted.<br/><br/>In addition to this, a method to interpolate between the signatures of new physics models is presented, which uses normalizing flows. The thesis then goes on to show the use of the technique in a search for a new scalar boson 𝑆 produced in association with a Higgs boson from a heavy new scalar 𝑋. There are also some contributions that interpolate between the event yields with Gaussian processes, and that show how we can use normalizing flows to construct a likelihood ratio-inspired observable.}}, author = {{Simpson, Nathan}}, isbn = {{978-91-8039-494-9}}, keywords = {{differentiable programming; machine learning; particle physics}}, language = {{eng}}, publisher = {{Lund University}}, school = {{Lund University}}, title = {{Data Analysis in High-Energy Physics as a Differentiable Program}}, url = {{http://dx.doi.org/10.5281/zenodo.7520315}}, doi = {{10.5281/zenodo.7520315}}, year = {{2023}}, }