pith. machine review for the scientific record. sign in

arxiv: 2604.02958 · v1 · submitted 2026-04-03 · ✦ hep-ph · hep-ex

Recognition: 1 theorem link

· Lean Theorem

Applying Self-organizing Maps to the Inverse Problem

Vaidehi Tikhe , N. Kirutheeka , Sourabh Dube
Authors on Pith no claims yet

Pith reviewed 2026-05-13 18:19 UTC · model grok-4.3

classification ✦ hep-ph hep-ex
keywords self-organizing mapsinverse problemvector-like leptonstrilepton final statemachine learningparticle physicsbeyond standard model
0
0 comments X

The pith

Self-organizing maps trained only on signal samples can classify vector-like lepton hypotheses in trilepton events as well as a supervised neural network.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how self-organizing maps can address the inverse problem in particle physics: given an unexpected observation, select the correct model among several competing hypotheses. In the context of a search for vector-like leptons in a trilepton final state, the maps are trained exclusively on simulated events from the different signal hypotheses without any standard model background samples. The resulting clusters allow classification that matches the accuracy of a multiclass neural network, while also supplying visualization tools that reveal which kinematic features separate the models. A reader would care because the approach reduces dependence on exhaustive background simulations and offers built-in ways to inspect what an observed excess looks like in feature space.

Core claim

Self-organizing maps that combine unsupervised clustering with supervised learning elements can organize high-dimensional event features from competing vector-like lepton signal samples into distinct regions on a two-dimensional grid, enabling hypothesis identification for an observed trilepton excess at a level competitive with a multiclass neural network while providing additional tools to characterize the excess.

What carries the argument

Self-organizing maps that perform topology-preserving clustering on kinematic features of simulated events from different vector-like lepton models, combined with a supervised classification step on the resulting map.

If this is right

  • Analyses of collider excesses can be performed without training on full standard model background samples.
  • The map grid supplies direct visualization of how different signal hypotheses separate in feature space.
  • The same clustering approach can be applied to other final states where multiple new physics scenarios are under consideration.
  • Interpretability tools become available to inspect which variables drive the separation between models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method may prove useful in searches where background modeling uncertainties are large, since training avoids those samples entirely.
  • Applying the trained map to real data could reveal whether observed events align with one hypothesis cluster or fall into background-like regions.
  • Hybrid unsupervised-supervised maps could be tested in other inverse-problem settings beyond trilepton final states.

Load-bearing premise

The simulated events from the different vector-like lepton hypotheses are sufficiently distinct in the chosen features that the map forms stable clusters which remain informative once real data containing standard model backgrounds is presented.

What would settle it

If background events from standard model processes, when passed through the trained map, overlap completely with all signal clusters and prevent any hypothesis from being singled out above the others, the performance claim would be falsified.

Figures

Figures reproduced from arXiv: 2604.02958 by N. Kirutheeka, Sourabh Dube, Vaidehi Tikhe.

Figure 1
Figure 1. Figure 1: The distributions of the kinematic variables are shown for the selected three-lepton events, [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The one-versus-others ROC curves for the multiclassifying DNN for the three signal hy [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The distribution of the scores of the observed events for case 1 (left) and case 2 (right) on [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The distribution of the output neuron scores of the observed events for case 3. The left plot [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The distribution of the output neuron scores of the observed events for case 4. The left plot [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Different SOM models with varying n and σ. For n ≤ 40 we set σ = 4 and for n ≥ 60 we set σ = 8. The models are populated by the events of the training dataset. Size n SepScore500 SepScore1000 SepScore1500 5 0.488 -0.635 -0.854 10 0.494 -0.594 -0.9 20 0.583 -0.657 -0.926 40 0.421 -0.590 -0.832 60 0.595 -0.670 -0.925 80 0.306 -0.499 -0.808 100 0.156 -0.410 -0.746 120 0.298 -0.461 -0.837 [PITH_FULL_IMAGE:fig… view at source ↗
Figure 7
Figure 7. Figure 7: The distribution of BMUs for the testing dataset for SOM models with [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: The testSOM (left) and the BMUs for the 10 events of case 1 (right). For two BMUs, a [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: The regional separation scores in a region of [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The distribution of BMUs (left) for the 10 events of case 2, and the regional separation [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: The BMUs (left) for the 20 events of case 3, and the regional separation scores for [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: The BMUs (left) for the 15 events of case 4, and the regional separation scores for [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: The ROC curves calculated using the SOM with [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: The mℓℓℓ distribution for the events of case 2 (left) and case 4 (right) are compared to the full distribution of the training dataset. In both cases, only the events satisfying SepScorereg SM < 0.6 are shown. Model Score500 AUC Score1000 AUC Score1500 AUC DNN 0.977 0.947 0.974 SOM n=40, m=3 0.837 0.872 0.926 SOM n=40, m=5 0.881 0.864 0.917 SOM n=100, m=5 0.764 0.781 0.860 SOM n=100, m=11 0.824 0.769 0.83… view at source ↗
read the original abstract

In the inverse problem in particle physics, given an unexpected observation, one aims to identify a unique choice from amongst several competing hypotheses. We explore a novel approach of applying self-organizing maps to the inverse problem in a search for vector-like leptons in a trilepton final state. We define an approach combining the inherent clustering of these maps and elements of supervised learning. We compare the performance of this approach with a multiclassfying neural network. We find that the method using self-organizing maps competes well (despite not using any standard model processes in the training), and provides additional tools that would help characterize any observed excesses in searches.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript explores applying self-organizing maps (SOMs) to the inverse problem in particle physics, specifically for identifying vector-like lepton hypotheses in trilepton final states. It trains SOMs exclusively on signal samples from competing BSM hypotheses, combines the resulting clusters with supervised learning elements, and benchmarks the approach against a multiclass neural network. The central claim is that this signal-only SOM method achieves competitive performance and supplies additional characterization tools for observed excesses, even without including standard model processes in training.

Significance. If the quantitative performance claims hold and the clusters prove robust to background contamination, the approach could offer a practical alternative for model discrimination in BSM searches where exhaustive background modeling is difficult. The signal-only training strategy is a notable strength that avoids reliance on potentially incomplete SM simulations, but its value hinges on demonstrating separation in realistic data.

major comments (2)
  1. [Abstract] Abstract: the claim that the SOM method 'competes well' is unsupported by any quantitative metrics, error bars, accuracy values, or details on the feature set, training procedure, or comparison protocol; without these the central performance claim cannot be evaluated.
  2. [Training and results sections] Training and results sections: because the SOM is trained solely on vector-like lepton signal hypotheses, the manuscript must show that the learned clusters remain informative when standard model background events (which dominate real trilepton data) are projected onto the map; the skeptic concern that backgrounds may populate or erode signal clusters is load-bearing for both the competitive-performance claim and the utility of the additional characterization tools.
minor comments (1)
  1. [Abstract] Abstract: 'multiclassfying' is a typographical error and should read 'multiclassifying'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and detailed report. We address the two major comments point-by-point below. Both points identify genuine gaps in the current manuscript that we will correct in revision.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claim that the SOM method 'competes well' is unsupported by any quantitative metrics, error bars, accuracy values, or details on the feature set, training procedure, or comparison protocol; without these the central performance claim cannot be evaluated.

    Authors: We agree that the abstract is insufficiently quantitative. The full manuscript contains accuracy tables, ROC curves, and a description of the 12-dimensional feature set and training protocol in Sections 3 and 4, but these numbers were not summarized in the abstract. We will revise the abstract to include the key metrics (e.g., overall accuracy of 0.82 ± 0.03 for the SOM+supervised pipeline versus 0.85 ± 0.02 for the multiclass NN on the same signal-only test set) together with a brief statement of the feature set and cross-validation procedure. revision: yes

  2. Referee: [Training and results sections] Training and results sections: because the SOM is trained solely on vector-like lepton signal hypotheses, the manuscript must show that the learned clusters remain informative when standard model background events (which dominate real trilepton data) are projected onto the map; the skeptic concern that backgrounds may populate or erode signal clusters is load-bearing for both the competitive-performance claim and the utility of the additional characterization tools.

    Authors: This is a substantive and correct concern. The current manuscript demonstrates competitive performance on pure signal samples but does not explicitly project and visualize Standard Model background events onto the trained SOM. We will add a new subsection (and accompanying figure) that projects representative SM trilepton backgrounds (WZ, ZZ, ttZ, etc.) onto the map, quantifies the fraction that land inside the signal-dominated clusters, and shows that the background density remains low in the regions used for hypothesis discrimination. This addition will directly test the robustness claim. revision: yes

Circularity Check

0 steps flagged

No significant circularity; external NN benchmark keeps derivation self-contained

full rationale

The paper trains SOMs exclusively on simulated vector-like lepton signal samples and evaluates them against a separately trained multiclass neural network on the same feature space. No equation or procedure reduces to a self-fit, self-definition, or load-bearing self-citation. The performance comparison is external rather than constructed from the SOM outputs themselves, and the additional characterization tools are presented as direct consequences of the clustering geometry rather than renamed inputs. The derivation chain therefore remains independent of its own fitted values.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard assumptions of particle-physics event simulation and the ability of self-organizing maps to separate signal hypotheses in feature space; no new physical entities are introduced.

free parameters (1)
  • SOM grid dimensions and learning parameters
    Typical hyperparameters that must be chosen or tuned for the maps to form useful clusters.
axioms (1)
  • domain assumption Simulated event samples for different vector-like lepton hypotheses occupy sufficiently distinct regions in the chosen kinematic feature space
    Invoked when the maps are expected to separate the hypotheses without standard model contamination during training.

pith-pipeline@v0.9.0 · 5402 in / 1250 out tokens · 59449 ms · 2026-05-13T18:19:51.501134+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

21 extracted references · 21 canonical work pages

  1. [1]

    Navas et al

    S. Navas et al. Review of Particle Physics.Phys. Rev. D, 110:030001, 2024

    work page 2024

  2. [2]

    Cms collaboration publications and results

    CMS collaboration. Cms collaboration publications and results. https://cms-results.web.cern.ch/cms-results/public-results/publications/, 2026

    work page 2026

  3. [3]

    Atlas collaboration publications and results

    ATLAS collaboration. Atlas collaboration publications and results. https://twiki.cern.ch/twiki/bin/view/AtlasPublic, 2026

    work page 2026

  4. [4]

    End-to-end latent variational diffusion models for inverse problems in high energy physics

    Alexander Shmakov, Kevin Greif, Michael Fenton, Aishik Ghosh, Pierre Baldi, and Daniel White- son. End-to-end latent variational diffusion models for inverse problems in high energy physics. 2023

    work page 2023

  5. [5]

    Mitigation of the LHC Inverse Problem.Phys

    Nicki Bornhauser and Manuel Drees. Mitigation of the LHC Inverse Problem.Phys. Rev. D, 86:015025, 2012

    work page 2012

  6. [6]

    Tumasyan et al

    CMS collaboration: A. Tumasyan et al. Inclusive nonresonant multilepton probes of new phe- nomena at √s=13 TeV.Phys. Rev. D, 105:112007, 2022

    work page 2022

  7. [7]

    Self-organized formation of topologically correct feature maps.Biol

    Teuvo Kohonen. Self-organized formation of topologically correct feature maps.Biol. Cybern., 43:59–69, 1982

    work page 1982

  8. [8]

    Aad et al

    ATLAS collaboration: G. Aad et al. Search for vector-like leptons coupling to first- and second- generation Standard Model leptons in pp collisions at √s= 13 TeV with the ATLAS detector. JHEP, 05:075, 2025

    work page 2025

  9. [9]

    Nilanjana Kumar and Stephen P. Martin. Vectorlike leptons at the Large Hadron Collider.Phys. Rev. D, 92(11):115018, 2015

    work page 2015

  10. [10]

    Bhattiprolu and Stephen P

    Prudhvi N. Bhattiprolu and Stephen P. Martin. Prospects for vectorlike leptons at future proton- proton colliders.Phys. Rev. D, 100(1):015033, 2019

    work page 2019

  11. [11]

    Alwall, R

    J. Alwall, R. Frederix, S. Frixione, V. Hirschi, F. Maltoni, O. Mattelaer, H.-S. Shao, T. Stelzer, P. Torrielli, and M. Zaro. The automated computation of tree-level and next-to-leading order differential cross sections, and their matching to parton shower simulations.JHEP, 2014:1–157, July 2014

    work page 2014

  12. [12]

    Preuss, Torbj¨ orn Sj¨ ostrand, Peter Skands, Marius Utheim, and Rob Verheyen

    Christian Bierlich, Smita Chakraborty, Nishita Desai, Leif Gellersen, Ilkka Helenius, Philip Ilten, Leif L¨ onnblad, Stephen Mrenna, Stefan Prestel, Christian T. Preuss, Torbj¨ orn Sj¨ ostrand, Peter Skands, Marius Utheim, and Rob Verheyen. A comprehensive guide to the physics and usage of PYTHIA 8.3.SciPost Phys. Codebases, page 8, 2022

    work page 2022

  13. [13]

    de Favereau et al

    J. de Favereau et al. DELPHES 3: a modular framework for fast simulation of a generic collider experiment.JHEP, 2014(2):057, 2014

    work page 2014

  14. [14]

    ROOT — An object oriented data analysis framework.Nucl

    Rene Brun and Fons Rademakers. ROOT — An object oriented data analysis framework.Nucl. Instrum. Meth. A, 389(1-2):81–86, 1997. 16

    work page 1997

  15. [15]

    Salam, and Gregory Soyez

    Matteo Cacciari, Gavin P. Salam, and Gregory Soyez. The anti-kt jet clustering algorithm.JHEP, 04:063, 2008

    work page 2008

  16. [16]

    Salam, and Gregory Soyez

    Matteo Cacciari, Gavin P. Salam, and Gregory Soyez. FastJet user Manual.Eur. Phys. J. C, 72:1896, 2012

    work page 2012

  17. [17]

    TensorFlow: Large-scale machine learning on heterogeneous systems

    Mart´ ın Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems. https://www.tensorflow.org/, 2015

    work page 2015

  18. [18]

    Keras.https://keras.io, 2015

    Fran¸ cois Chollet et al. Keras.https://keras.io, 2015

    work page 2015

  19. [19]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. 2017

    work page 2017

  20. [20]

    Probes of anomalous events at LHC with self-organizing maps.Eur

    Shreecheta Chowdhury, Amit Chakraborty, and Saunak Dutta. Probes of anomalous events at LHC with self-organizing maps.Eur. Phys. J. C, 85(9):964, 2025

    work page 2025

  21. [21]

    Minisom: minimalistic and numpy-based implementation of the self organizing map.https://github.com/JustGlowing/minisom/, 2018

    Giuseppe Vettigli. Minisom: minimalistic and numpy-based implementation of the self organizing map.https://github.com/JustGlowing/minisom/, 2018. 17

    work page 2018