SciPost Submission Page
SPANet: Generalized Permutationless Set Assignment for Particle Physics using Symmetry Preserving Attention
by Alexander Shmakov, Michael James Fenton, Ta-Wei Ho, Shih-Chieh Hsu, Daniel Whiteson, Pierre Baldi
This is not the latest submitted version.
This Submission thread is now published as
Submission summary
| Authors (as registered SciPost users): | Michael James Fenton |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2106.03898v2 (pdf) |
| Code repository: | https://github.com/Alexanders101/SPANet |
| Data repository: | http://mlphysics.ics.uci.edu/data/2021_ttbar/ |
| Date submitted: | 2021-11-03 01:42 |
| Submitted by: | Fenton, Michael James |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Experimental, Computational |
Abstract
The creation of unstable heavy particles at the Large Hadron Collider is the most direct way to address some of the deepest open questions in physics. Collisions typically produce variable-size sets of observed particles which have inherent ambiguities complicating the assignment of observed particles to the decay products of the heavy particles. Current strategies for tackling these challenges in the physics community ignore the physical symmetries of the decay products and consider all possible assignment permutations and do not scale to complex configurations. Attention based deep learning methods for sequence modelling have achieved state-of-the-art performance in natural language processing, but they lack built-in mechanisms to deal with the unique symmetries found in physical set-assignment problems. We introduce a novel method for constructing symmetry-preserving attention networks which reflect the problem's natural invariances to efficiently find assignments without evaluating all permutations. This general approach is applicable to arbitrarily complex configurations and significantly outperforms current methods, improving reconstruction efficiency between 19\% - 35\% on typical benchmark problems while decreasing inference time by two to five orders of magnitude on the most complex events, making many important and previously intractable cases tractable. A full code repository containing a general library, the specific configuration used, and a complete dataset release, are avaiable at https://github.com/Alexanders101/SPANet
Current status:
Reports on this Submission
Report #1 by Anonymous (Referee 1) on 2021-12-10 (Invited Report)
- Cite as: Anonymous, Report on arXiv:2106.03898v2, delivered 2021-12-10, doi: 10.21468/SciPost.Report.4020
Strengths
1: Proposes an innovative ML-based method for a set-based assignment problem in particle physics (jet-parton assignment) taking into account permutation symmetries of the system.
2: Provides results on three benchmark problems of increasing complexity together with an analysis of runtime properties. The results show a clear improvement, especially for the most complex problem considered (tttt)
3) As e.g. precision measurements of the Higgs sector increasingly tackle complex final states, the proposed method to resolve the parton assignment is timely and very well motivated. The provided implementation should experimental deployment.
Weaknesses
1) In the context of baseline methods, a discussion of the Matrix Element method would be useful. For example, there exist results by the CMS collaboraton using a matrix element approach for the all-hadronic ttH final state (1803.06986,). The sentence "there exists no study of the ttH¯ process in which all partons lead to jets which attempts a full event reconstruction" should be amended accordingly.
2) Table 2: Why is the matching efficiency using the chi2 method lower for complete events than for all events? Can this be understood in more detail?
Report
Parton assignment is an important aspect of event classification in particle physics, and required for many analyses of complex final states. Especially in future data taking periods of the LHC/HL-LHC, measurements of such topologies will become increasingly relevant.
The paper makes important progress on this challenges by proposing a novel and well-principled machine learning approach to efficiently assign permutations.
We comment on some minor issues above, but once these are resolved fully recommend the manuscript for publication in SciPost.
Requested changes
1) Include discussion of MEM (see above)
2) Consider moving section 6 (Codebase) to an Appendix
Author: Michael James Fenton on 2021-12-23 [id 2052]
(in reply to Report 1 on 2021-12-10)We have edited the baseline discussion section as follows;
"To our knowledge, there exists no study of the \ttH{} process in which all partons lead to jets which attempts a full event reconstruction, and we are further not aware of any analysis of all-jet \tttt{} at all."
Changed to
"To our knowledge, the only study of the \ttH{} process in which all partons lead to jets which attempts a full event reconstruction is \cite{CMS:2018sah}, which uses a matrix element method (MEM) to simultaneously reconstruct the event and separate signal and background. Unfortunately, this result does not report any results for the reconstruction efficiency, and the main purpose of the MEM appears to be the signal and background separation rather than the event reconstruction. We are further not aware of any analysis of all-jet \tttt{} at all."
2) Table 2: Why is the matching efficiency using the chi2 method lower for complete events than for all events? Can this be understood in more detail?
This appears to be a consequence of the cuts and data distributions within each cut. If you look at the event fractions, the “Complete” events are mostly dominated by high jet-multiplicity events, i.e. more difficult ones. This makes intuitive sense since higher multiplicity events are more likely to be successfully truth matched. The chi^2 performance degrades rapidly vs NJets, reducing efficiency numbers on these more challenging cuts.
It is the preference of the authors to keep section 6 in the main body rather than moving to an appendix, but we are open to this change if the editor prefers it. We have kept it in the main body for this revision.