arxiv: 2604.05290 · v1 · submitted 2026-04-07 · 🌌 astro-ph.CO

Recognition: no theorem link

Explaining Neural Networks on the Sky: Machine Learning Interpretability for Cosmic Microwave Background Maps

Indira Ocampo , Guadalupe Ca\~nas-Herrera

Authors on Pith no claims yet

Pith reviewed 2026-05-10 19:34 UTC · model grok-4.3

classification 🌌 astro-ph.CO

keywords CMB mapsneural networksSHAP interpretabilityprimordial featuresLambda CDMcosmological model selectionmap-level analysisinflationary signals

0 comments

The pith

Neural networks trained directly on CMB maps identify subtle primordial features by retaining full spatial information that power spectra discard.

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

The paper develops neural networks that classify simulated CMB temperature and polarization maps to separate the standard Lambda CDM model from versions containing extra primordial features. Unlike conventional analyses that compress data into angular power spectra, this approach works with the complete map-level variations across the sky. It then uses SHAP explanations to reveal which specific sky patches and angular scales most influence the model's decisions. A reader would care because the method offers a way to spot faint early-universe signals that summary statistics might miss, while releasing a public pipeline for simulation, training, and interpretation.

Core claim

By training neural networks directly on simulated Cosmic Microwave Background maps rather than on compressed angular power spectra, the framework retains the full spatial information of temperature and polarization anisotropies. This enables the identification of subtle signatures of primordial features beyond the standard Lambda CDM model. SHAP attributions then pinpoint the sky regions and scales that contribute most to distinguishing between models, providing a proof-of-concept for interpretable machine learning applied to CMB data.

What carries the argument

Map-level neural network classifier with principal component analysis preprocessing and post-hoc SHAP attributions to interpret spatial contributions to model classification.

If this is right

Retains the full spatial information of temperature and polarisation anisotropies instead of using only averaged power spectra.
Enables identification of subtle signatures of primordial features in the maps.
SHAP analysis reveals which regions of the sky and which scales drive the distinction between Lambda CDM and feature models.
Serves as a proof-of-concept for uncovering higher-order information in CMB data through interpretable machine learning.
Provides public access to the complete pipeline for map generation, network training, and interpretability analysis.

Where Pith is reading between the lines

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

If the approach succeeds on real data, it could be tested on higher-resolution maps from next-generation CMB experiments to search for localized inflationary signals.
The open pipeline allows direct comparison with traditional power-spectrum methods on the same simulated datasets to quantify the gain from spatial information.
Combining these map-level attributions with cross-correlations to other probes such as large-scale structure might tighten constraints on early-universe physics.

Load-bearing premise

Neural networks trained only on simulated maps will generalize to real observations and SHAP attributions will reflect genuine physical distinctions rather than simulation or training artifacts.

What would settle it

Applying the trained classifier to actual Planck CMB maps and finding that SHAP-highlighted regions show no consistent alignment with expected physical features or that classification performance collapses compared to simulation results.

read the original abstract

We present a framework for cosmological model selection using Neural Networks (NNs) trained directly on simulated Cosmic Microwave Background (CMB) temperature and polarisation maps. By operating at the map level rather than on compressed angular power spectra, our approach retains the full spatial information of temperature and polarisation anisotropies, enabling the identification of subtle signatures of primordial features beyond the standard $\Lambda$CDM model. We describe the generation of Planck-like CMB maps, and the hybrid architecture that combines principal component analysis and neural networks, optimised for classification tasks. To understand how the classifier reaches its decisions, we apply Shapley Additive exPlanations (SHAP) as a post-hoc interpretability tool, identifying which regions of the sky and which scales contribute most to the distinction between $\Lambda$CDM and feature models. This work serves as a follow-up to previous analyses at the level of summary statistics and as a proof-of-concept for using interpretable machine learning to uncover higher-order information in CMB data, with the potential to enhance the detection of nontrivial inflationary signals and improve cosmological model discrimination. Results for model classification performance, calibration, and interpretability are presented as a placeholder for the full analysis. In addition, we introduce the Open Science project, providing public access to the full pipeline for simulation, training, and interpretability of CMB map-based neural networks.

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.

Desk Editor's Note private letter to a colleague

This manuscript describes an intended framework for neural network classification of CMB maps using SHAP interpretability, but it contains no actual results or metrics.

read the letter

The key point here is that this is a methods proposal with placeholder results, so the claim that map-level neural networks can identify primordial features better than spectrum-based approaches stays untested. The paper sets out a clear plan: generate simulated CMB maps that include both temperature and polarization, feed them into a hybrid model that first reduces dimensions with principal component analysis and then classifies with a neural network, and finally use SHAP values to highlight which sky regions and scales drive the classification between Lambda CDM and feature models. It positions this as a follow-up to earlier summary-statistic work and emphasizes keeping the full spatial information instead of compressing it early. The commitment to an open science project with public code for the entire pipeline is a practical contribution that others could build on. Where it falls short is the lack of any supporting evidence. The abstract notes that classification performance, calibration, and interpretability results are placeholders for the full analysis. Without numbers on accuracy, without confusion matrices or SHAP plots, and without direct comparisons to power-spectrum methods, there is no way to check if the approach actually extracts higher-order information or if the attributions are meaningful. The assumption that the network will generalize from simulations to real Planck data also goes unexamined in the current text. This kind of paper would interest researchers who are already exploring machine learning for cosmological data analysis and who might want to try the released code. A reader looking for new constraints on inflation or concrete improvements in model selection would not find them here. The thinking behind the framework is coherent and engages honestly with the limitations of summary statistics, but the work is not yet at the stage where it can be evaluated on its empirical merits. I would not bring this to a reading group unless the group focuses on methodological proposals. I would not cite it in my own work until results appear. It does not seem ready for peer review; the authors should complete the analysis and include the metrics before submitting.

Referee Report

1 major / 0 minor

Summary. The manuscript proposes a framework for cosmological model selection by training neural networks directly on simulated Planck-like CMB temperature and polarization maps rather than on angular power spectra. It describes map generation, a hybrid PCA+NN architecture for classification between ΛCDM and primordial-feature models, and the application of SHAP values to identify spatially localized contributions to the decisions. The work is positioned as a proof-of-concept and follow-up to summary-statistic analyses, with an accompanying open-science pipeline release; however, all quantitative results on classification performance, calibration, and interpretability are explicitly noted as placeholders.

Significance. If the promised results were to demonstrate that the map-level hybrid model extracts higher-order spatial information beyond what is captured by power spectra, and that the resulting SHAP attributions align with physically motivated distinctions, the approach could meaningfully extend model-discrimination capabilities for inflationary features. The open-science release of the full simulation-training-interpretability pipeline is a clear positive contribution that would facilitate reproducibility and community follow-up regardless of the specific performance numbers.

major comments (1)

[Abstract / Results] Abstract and results section: the central claim that operating at the map level 'retains the full spatial information ... enabling the identification of subtle signatures of primordial features' is presented without any supporting metrics, confusion matrices, SHAP maps, or baseline comparisons to spectrum-based classifiers. The text states that 'Results for model classification performance, calibration, and interpretability are presented as a placeholder for the full analysis,' leaving the load-bearing assertion untested.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their constructive report and for recognizing the potential value of the map-level framework and the open-science pipeline release. We agree that the current manuscript version leaves the central claims unsupported due to the use of placeholders, and we will revise accordingly to provide the full quantitative analysis.

read point-by-point responses

Referee: [Abstract / Results] Abstract and results section: the central claim that operating at the map level 'retains the full spatial information ... enabling the identification of subtle signatures of primordial features' is presented without any supporting metrics, confusion matrices, SHAP maps, or baseline comparisons to spectrum-based classifiers. The text states that 'Results for model classification performance, calibration, and interpretability are presented as a placeholder for the full analysis,' leaving the load-bearing assertion untested.

Authors: We acknowledge that the manuscript as submitted relies on placeholders for all quantitative results on classification performance, calibration, and interpretability, which means the key assertions about retaining full spatial information and identifying subtle primordial feature signatures lack empirical support in the current draft. This version was prepared as a methodological proof-of-concept describing the simulation pipeline, hybrid PCA+NN architecture, and SHAP interpretability approach, with the explicit intention of completing the numerical evaluation prior to final publication. In the revised manuscript we will replace the placeholders with the complete results, including classification metrics (accuracy, precision, recall, AUC), confusion matrices, calibration plots, SHAP attribution maps showing spatially localized contributions, and direct baseline comparisons against classifiers trained on angular power spectra. The abstract and results section will be updated to reflect these findings and to ensure all claims are properly substantiated. revision: yes

Circularity Check

0 steps flagged

No circularity; paper is a methodological proposal with placeholder results and no derivations

full rationale

The manuscript presents a framework for NN-based CMB map classification and SHAP interpretability but contains no equations, fitted parameters, derivations, or quantitative results. All performance claims are explicitly labeled as placeholders for future analysis, and the central claim about retaining full spatial information is a descriptive statement of the method rather than a derived prediction. No self-citations are load-bearing for any result, and the work is self-contained as a proof-of-concept description without internal reduction to inputs.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The description relies on standard assumptions for generating Planck-like CMB simulations from LambdaCDM and feature models; no free parameters, new entities, or ad-hoc axioms are introduced in the abstract.

axioms (1)

domain assumption Standard cosmological assumptions underlying the generation of simulated CMB maps from LambdaCDM and primordial feature models
The framework presupposes that these simulations accurately capture the relevant differences between models.

pith-pipeline@v0.9.0 · 5545 in / 1319 out tokens · 65425 ms · 2026-05-10T19:34:36.458493+00:00 · methodology

discussion (0)

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Modifications of CMB Temperature and Polarization Quadrupole Signals in Thurston Spacetimes
gr-qc 2026-05 unverdicted novelty 6.0

Thurston spacetimes generate distinct evolving temperature and polarization patterns in the CMB that can be tracked via Stokes parameters and potentially isolated per geometry.

Reference graph

Works this paper leans on

43 extracted references · 5 canonical work pages · cited by 1 Pith paper · 1 internal anchor

[1]

Chang et al.,Snowmass2021 Cosmic Frontier: Cosmic Microwave Background Measurements White Paper,2203.07638

Clarence L Chang et al. Snowmass2021 cosmic frontier: Cosmic microwave background measurements white paper.arXiv preprint arXiv:2203.07638, 2022. – 14 –

work page arXiv 2022
[2]

TASI Lectures on Inflation

Daniel Baumann. Tasi lectures on inflation.arXiv preprint arXiv:0907.5424, 2009

work page Pith review arXiv 2009
[3]

Dynamical dark energy in light of the desi dr2 baryonic acoustic oscillations measurements.arXiv preprint arXiv:2504.06118, 2025

Gan Gu, Xiaoma Wang, Yuting Wang, Gong-Bo Zhao, Levon Pogosian, Kazuya Koyama, John A Peacock, Zheng Cai, Jorge L Cervantes-Cota, Ruiyang Zhao, et al. Dynamical dark energy in light of the desi dr2 baryonic acoustic oscillations measurements.arXiv preprint arXiv:2504.06118, 2025

work page arXiv 2025
[4]

Bayesian analysis of inflationary features in planck and sdss data.Physical Review D, 94(2):023526, 2016

Micol Benetti and Jailson S Alcaniz. Bayesian analysis of inflationary features in planck and sdss data.Physical Review D, 94(2):023526, 2016

2016
[5]

Neural networks for cosmological model selection and feature importance using cosmic microwave background data.Journal of Cosmology and Astroparticle Physics, 2025(02):004, 2025

I Ocampo, G Cañas-Herrera, and S Nesseris. Neural networks for cosmological model selection and feature importance using cosmic microwave background data.Journal of Cosmology and Astroparticle Physics, 2025(02):004, 2025

2025
[6]

Investigating the constraints on primordial features with future cosmic microwave background and galaxy surveys.Journal of Cosmology and Astroparticle Physics, 2022(09):024, 2022

Debabrata Chandra and Supratik Pal. Investigating the constraints on primordial features with future cosmic microwave background and galaxy surveys.Journal of Cosmology and Astroparticle Physics, 2022(09):024, 2022

2022
[7]

Enhancing cmb map reconstruction and power spectrum estimation with convolutional neural networks.Journal of Cosmology and Astroparticle Physics, 2024(04):041, 2024

Belén Costanza, Claudia G Scóccola, and Matías Zaldarriaga. Enhancing cmb map reconstruction and power spectrum estimation with convolutional neural networks.Journal of Cosmology and Astroparticle Physics, 2024(04):041, 2024

2024
[8]

Physical Review D, 111(8):083537, 2025

Davide Piras, Laura Herold, Luisa Lucie-Smith, and Eiichiro Komatsu.λcdm and early dark energy in latent space: A data-driven parametrization of the cmb temperature power spectrum. Physical Review D, 111(8):083537, 2025

2025
[9]

The importance of being interpretable: Toward an understandable machine learning encoder for galaxy cluster cosmology.The Astrophysical Journal, 926(1):45, 2022

Michelle Ntampaka and Alexey Vikhlinin. The importance of being interpretable: Toward an understandable machine learning encoder for galaxy cluster cosmology.The Astrophysical Journal, 926(1):45, 2022

2022
[10]

Enhancing cosmological model selection with interpretable machine learning.Physical Review Letters, 134(4):041002, 2025

Indira Ocampo, George Alestas, Savvas Nesseris, and Domenico Sapone. Enhancing cosmological model selection with interpretable machine learning.Physical Review Letters, 134(4):041002, 2025

2025
[11]

Lundberg and Su-In Lee

Scott M. Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors,Advances in Neural Information Processing Systems 30, pages 4765–4774. Curran Associates, Inc., 2017

2017
[12]

Searching for local features in primordial power spectrum using genetic algorithms.arXiv preprint arXiv:2308.04940, 2023

Kushal Lodha, Lucas Pinol, Savvas Nesseris, Arman Shafieloo, Wuhyun Sohn, and Matteo Fasiello. Searching for local features in primordial power spectrum using genetic algorithms.arXiv preprint arXiv:2308.04940, 2023

work page arXiv 2023
[13]

Machine learning cosmic inflation.Physical Review D, 108(4):043509, 2023

Ahana Kamerkar, Savvas Nesseris, and Lucas Pinol. Machine learning cosmic inflation.Physical Review D, 108(4):043509, 2023

2023
[14]

Jens Chluba, Jan Hamann, and Subodh P. Patil. Features and new physical scales in primordial observables: Theory and observation.International Journal of Modern Physics D, 24(10):1530023, June 2015

2015
[15]

Ballardini et al

M. Ballardini et al. Euclid: The search for primordial features.Astron. Astrophys., 683:A220, 2024

2024
[16]

Akrami et al

Y. Akrami et al. Planck 2018 results. X. Constraints on inflation.Astron. Astrophys., 641:A10, 2020

2018
[17]

Cambridge University Press, 2020

Ruth Durrer.The cosmic microwave background. Cambridge University Press, 2020

2020
[18]

A cmb polarization primer.New Astronomy, 2(4):323–344, 1997

Wayne Hu and Martin White. A cmb polarization primer.New Astronomy, 2(4):323–344, 1997

1997
[19]

Górski, Eric Hivon, Anthony J

Krzysztof M. Górski, Eric Hivon, Anthony J. Banday, Benjamin D. Wandelt, Frode K. Hansen, Martin Reinecke, and Matthias Bartelmann. Healpix: A framework for high-resolution discretization and fast analysis of data distributed on the sphere.The Astrophysical Journal, 622(2):759–771, 2005

2005
[20]

Planck 2018 results-iii

Nabila Aghanim, Yashar Akrami, Mark Ashdown, J Aumont, Carlo Baccigalupi, M Ballardini, Anthony J Banday, RB Barreiro, Nicola Bartolo, S Basak, et al. Planck 2018 results-iii. high frequency instrument data processing and frequency maps.Astronomy & Astrophysics, 641:A3, 2020. – 15 –

2018
[21]

Cosmic microwave background polarization.Annals of Physics, 246(1):49–85, 1996

Arthur Kosowsky. Cosmic microwave background polarization.Annals of Physics, 246(1):49–85, 1996

1996
[22]

All-sky analysis of polarization in the microwave background

Matias Zaldarriaga and Uroš Seljak. All-sky analysis of polarization in the microwave background. Physical Review D, 55(4):1830–1840, 1997

1997
[23]

Statistics of cosmic microwave background polarization.Physical Review D, 55(12):7368–7388, 1997

Marc Kamionkowski, Arthur Kosowsky, and Albert Stebbins. Statistics of cosmic microwave background polarization.Physical Review D, 55(12):7368–7388, 1997

1997
[24]

The polarization of the cosmic microwave background due to primordial gravitational waves

Brian G Keating, Alexander G Polnarev, Nathan J Miller, and Deepak Baskaran. The polarization of the cosmic microwave background due to primordial gravitational waves. International Journal of Modern Physics A, 21(12):2459–2479, 2006

2006
[25]

Second season quiet observations: Measurements of the cosmic microwave background polarization power spectrum at 95 ghz.The Astrophysical Journal, 760(2):145, 2012

Quiet Collaboration, D Araujo, C Bischoff, A Brizius, I Buder, Y Chinone, K Cleary, RN Dumoulin, A Kusaka, R Monsalve, et al. Second season quiet observations: Measurements of the cosmic microwave background polarization power spectrum at 95 ghz.The Astrophysical Journal, 760(2):145, 2012

2012
[26]

Methods for cmb map analysis

Raelyn M Sullivan, Lukas Tobias Hergt, and Douglas Scott. Methods for cmb map analysis. Research Notes of the AAS, 9(2):43, 2025

2025
[27]

Planck 2018 results

Planck Collaboration. Planck 2018 results. i. overview and the cosmological legacy of planck. Astronomy & Astrophysics, 641:A1, 2020

2018
[28]

Camb: Code for anisotropies in the microwave background

Antony Lewis and Anthony Challinor. Camb: Code for anisotropies in the microwave background. Astrophysics source code library, pages ascl–1102, 2011

2011
[29]

Efficient computation of cosmic microwave background anisotropies inclosed friedmann-robertson-walker models.The Astrophysical Journal, 538(2):473, 2000

Antony Lewis, Anthony Challinor, and Anthony Lasenby. Efficient computation of cosmic microwave background anisotropies inclosed friedmann-robertson-walker models.The Astrophysical Journal, 538(2):473, 2000

2000
[30]

Mass reconstruction with cosmic microwave background polarization.Physical Review D, 69(4):043004, 2004

Wayne Hu and Takemi Okamoto. Mass reconstruction with cosmic microwave background polarization.Physical Review D, 69(4):043004, 2004

2004
[31]

Planck 2018 results

Planck Collaboration. Planck 2018 results. iv. diffuse component separation.Astronomy & Astrophysics, 641:A4, 2020

2018
[32]

Gorski, C

Eric Hivon, Krzysztof M. Gorski, C. B. Netterfield, B. P. Crill, Simon Prunet, and Frode Hansen. Master of the cmb anisotropy power spectrum: A fast method for statistical analysis of large and complex cmb data sets.The Astrophysical Journal, 567:2–17, 2002

2002
[33]

Healpix: A framework for high-resolution discretization and fast analysis of data distributed on the sphere.The Astrophysical Journal, 622(2):759, 2005

Krzysztof M Gorski, Eric Hivon, Anthony J Banday, Benjamin D Wandelt, Frode K Hansen, Mstvos Reinecke, and Matthia Bartelmann. Healpix: A framework for high-resolution discretization and fast analysis of data distributed on the sphere.The Astrophysical Journal, 622(2):759, 2005

2005
[34]

Impact of data normalization on deep neural network for time series forecasting

Samit Bhanja and Abhishek Das. Impact of data normalization on deep neural network for time series forecasting. InProceedings of conference on Advancement in Computation, Communication and Electronics Paradigm (ACCEP-2019), page 27, 2019

2019
[35]

Recovering the cmb signal with machine learning.The Astrophysical Journal Supplement Series, 260(1):13, 2022

Guo-Jian Wang, Hong-Liang Shi, Ye-Peng Yan, Jun-Qing Xia, Yan-Yun Zhao, Si-Yu Li, and Jun-Feng Li. Recovering the cmb signal with machine learning.The Astrophysical Journal Supplement Series, 260(1):13, 2022

2022
[36]

Pca based kernel initialization for convolutional neural networks

Yifeng Wang, Yuxi Rong, Hongyue Pan, Ke Liu, Yang Hu, Fangmin Wu, Wei Peng, Xingsi Xue, and Junfeng Chen. Pca based kernel initialization for convolutional neural networks. In International Conference on Data Mining and Big Data, pages 71–82. Springer, 2020

2020
[37]

Convolutional neural network based on principal component analysis initialization for image classification

Xu-Die Ren, Hao-Nan Guo, Guan-Chen He, Xu Xu, Chong Di, and Sheng-Hong Li. Convolutional neural network based on principal component analysis initialization for image classification. In 2016 IEEE first international conference on data science in cyberspace (DSC), pages 329–334. IEEE, 2016

2016
[38]

George Efstathiou. Principal-component analysis of the cosmic microwave background anisotropies: revealing the tensor degeneracy.Monthly Notices of the Royal Astronomical Society, 332(1):193–198, 2002

2002
[39]

Vinod Nair and Geoffrey E. Hinton. Rectified linear units improve restricted boltzmann machines. InProceedings of the 27th International Conference on Machine Learning (ICML), 2010. – 16 –

2010
[40]

MIT Press, 2016

Ian Goodfellow, Yoshua Bengio, and Aaron Courville.Deep Learning. MIT Press, 2016

2016
[41]

Adam: A Method for Stochastic Optimization

Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization.arXiv preprint arXiv:1412.6980, 2014

work page internal anchor Pith review Pith/arXiv arXiv 2014
[42]

An introduction to roc analysis.Pattern Recognition Letters, 27(8):861–874, 2006

Tom Fawcett. An introduction to roc analysis.Pattern Recognition Letters, 27(8):861–874, 2006

2006
[43]

From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020

Scott M Lundberg, Gabriel Erion, Hugh Chen, Alex DeGrave, Jordan M Prutkin, Bala Nair, Ronit Katz, Jonathan Himmelfarb, Nisha Bansal, and Su-In Lee. From local explanations to global understanding with explainable ai for trees.Nature machine intelligence, 2(1):56–67, 2020. – 17 – Appendix A: NN Compilation In Figure 7 we can see how the CNN compilation ti...

2020