pith. sign in

arxiv: 2606.08826 · v1 · pith:TLIQGKMXnew · submitted 2026-06-07 · 💻 cs.CV · astro-ph.GA

Classifying galaxies in the Galaxy10 DECals dataset using Inception and Residual CNNs

Lanz Anthonee A. Lagman , Prospero C. Naval Jr , Reinabelle C. Reyes This is my paper

Pith reviewed 2026-06-27 18:38 UTC · model grok-4.3

classification 💻 cs.CV astro-ph.GA
keywords galaxy classificationconvolutional neural networksResNetInceptionGalaxy10 DECalsimage augmentationgalaxy morphologydeep learning
0
0 comments X

The pith

ResNet101 and InceptionV4 both classify galaxies at about 90 percent accuracy, with ResNet101 performing better.

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

This paper tests ResNet101 and InceptionV4 convolutional neural networks on a version of the Galaxy10 DECals dataset where the number of images per galaxy class has been adjusted and spatial augmentations applied. The goal is to see if these efficient deep learning architectures can handle the growing volume of galaxy image data from future surveys. Both models reach roughly 90 percent accuracy, similar to other reported results, though ResNet101 shows better overall metrics. A sympathetic reader would care because manual classification cannot scale to the expected data flood, so reliable automated methods are needed. The work suggests these networks can form the base for such automated pipelines.

Core claim

In this work, we analyze the performance of ResNet101 and InceptionV4 on a spatially-augmented Galaxy10 DECals dataset. Retaining the ten-class classification of galaxies, we modify the image count of each class. We find that ResNet101 and InceptionV4 models achieved accuracies of ~ 90%, comparable with reported performance in the literature. In terms of performance metrics, ResNet101 is superior to InceptionV4. Our results indicate that either of these CNN architectures could serve as a robust foundation for specialized pipelines for classification of galaxy images from upcoming surveys.

What carries the argument

Comparison between ResNet101 residual connections and InceptionV4 parallel inception modules on the class-count-modified Galaxy10 DECals dataset for ten-class galaxy morphology classification.

If this is right

  • ResNet101 delivers superior performance metrics compared to InceptionV4.
  • Both models attain accuracies around 90 percent that match existing literature.
  • Either network can serve as a starting point for building classification systems for new galaxy surveys.
  • The approach of adjusting class image counts and applying spatial augmentation supports effective model training.

Where Pith is reading between the lines

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

  • If the balancing step proves unnecessary, simpler training on original distributions might suffice for similar performance.
  • These models could be adapted for real-time classification in upcoming survey pipelines if computational costs remain low.
  • Cross-validation on datasets from other telescopes would test whether the reported superiority generalizes beyond DECals data.

Load-bearing premise

Altering the image count per class in the Galaxy10 DECals dataset, combined with spatial augmentation, produces a training distribution whose performance metrics generalize to unaltered survey data without introducing class-balance artifacts that favor one architecture over the other.

What would settle it

Running the trained models on the original Galaxy10 DECals dataset without class count modification or augmentation and observing either substantially lower accuracy or InceptionV4 outperforming ResNet101.

Figures

Figures reproduced from arXiv: 2606.08826 by Lanz Anthonee A. Lagman, Prospero C. Naval Jr, Reinabelle C. Reyes.

Figure 1
Figure 1. Figure 1: Galaxy10 DECals dataset classes and respective number of images per class. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Loss curves from training and validation of the selected CNN architectures. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Normalized confusion matrices for ResNet101 and InceptionV4, showing accuracies per class in percent [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
read the original abstract

Image data regarding galactic morphology is expected to increase both in quantity and quality for the next foreseeable years; thus it is important to explore which deep learning architectures adapted for image classification tasks are cost-effective. Residual and Inception networks are ideal for exploring classification convolutional neural networks (CNNs) due to their computational efficiency, achieved through techniques such as residual connections and parallelized inception modules, enabling deeper networks without excessively increasing computational complexity. In this work, we analyze the performance of ResNet101 and InceptionV4 on a spatially-augmented Galaxy10 DECals dataset. Retaining the ten-class classification of galaxies, we modify the image count of each class. We find that ResNet101 and InceptionV4 models achieved accuracies of $\sim$ 90%, comparable with reported performance in the literature. In terms of performance metrics, ResNet101 is superior to InceptionV4. Our results indicate that either of these CNN architectures could serve as a robust foundation for specialized pipelines for classification of galaxy images from upcoming surveys.

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

3 major / 2 minor

Summary. The paper evaluates ResNet101 and InceptionV4 CNNs on the Galaxy10 DECals dataset after modifying per-class image counts and applying spatial augmentation. It reports ~90% classification accuracy for both models (comparable to literature) and claims ResNet101 is superior based on performance metrics, concluding that either architecture could foundation specialized pipelines for future galaxy surveys.

Significance. If the empirical results and superiority claim can be verified with full methodological details and ablations, the work would offer a practical benchmark for applying established residual and inception architectures to large-scale astronomical imaging, supporting cost-effective classification pipelines for upcoming surveys. The purely empirical nature provides no parameter-free derivations or machine-checked proofs.

major comments (3)
  1. [Abstract / Dataset section] Abstract and dataset description: the central superiority claim for ResNet101 rests on a training distribution created by explicitly modifying image counts per class before augmentation; no ablation is reported that tests whether the performance gap survives on the unaltered original class distribution or a held-out test set drawn from the survey's native balance.
  2. [Experimental setup / Results] Experimental setup / results: accuracy figures of ~90% and the ResNet101 vs. InceptionV4 comparison are stated without any training protocol, data-split details, hyperparameter settings, statistical tests, or error bars, rendering the performance comparison unverifiable from the given text.
  3. [Results] Results section: the claim that either architecture 'could serve as a robust foundation' for specialized pipelines is load-bearing on the reported metrics generalizing beyond the authors' chosen per-class counts, yet no sensitivity analysis to class-balance artifacts is provided.
minor comments (2)
  1. [Methods] Notation for model variants (ResNet101, InceptionV4) should be defined consistently with any cited literature implementations.
  2. [Figures / Tables] Figure captions and table headers lack explicit mention of whether metrics are computed on validation or test splits.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below, indicating revisions where appropriate to enhance methodological transparency and address concerns about class distribution effects.

read point-by-point responses

  1. Referee: [Abstract / Dataset section] Abstract and dataset description: the central superiority claim for ResNet101 rests on a training distribution created by explicitly modifying image counts per class before augmentation; no ablation is reported that tests whether the performance gap survives on the unaltered original class distribution or a held-out test set drawn from the survey's native balance.

    Authors: The per-class image counts were modified to mitigate the severe imbalance present in the original Galaxy10 DECals dataset, a standard approach for training stable classifiers. The test set was held out after this modification and augmentation. We did not perform the requested ablation on the unaltered native distribution. In revision we will report the exact original and modified counts, clarify that the superiority claim applies to the balanced setting, and add an explicit limitations paragraph noting the absence of native-balance testing. revision: yes

  2. Referee: [Experimental setup / Results] Experimental setup / results: accuracy figures of ~90% and the ResNet101 vs. InceptionV4 comparison are stated without any training protocol, data-split details, hyperparameter settings, statistical tests, or error bars, rendering the performance comparison unverifiable from the given text.

    Authors: We agree that the submitted text omitted these reproducibility details. The revised manuscript will expand the experimental setup section to specify the train/validation/test split ratios, optimizer, learning-rate schedule, batch size, epoch count, augmentation parameters, and any repeated runs with error bars or statistical tests. revision: yes

  3. Referee: [Results] Results section: the claim that either architecture 'could serve as a robust foundation' for specialized pipelines is load-bearing on the reported metrics generalizing beyond the authors' chosen per-class counts, yet no sensitivity analysis to class-balance artifacts is provided.

    Authors: The claim is qualified to the balanced, augmented distribution used in the experiments. We will add a short sensitivity discussion in the results section that examines how performance varies with different class-count ratios and will note the dependence on the chosen balancing strategy as a limitation. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical evaluation of off-the-shelf CNNs

full rationale

The paper reports accuracies from training ResNet101 and InceptionV4 on a modified Galaxy10 DECals dataset (altered per-class counts + spatial augmentation). No equations, derivations, fitted parameters renamed as predictions, or self-citation chains appear in the abstract or described content. The central claims are direct empirical measurements on the authors' chosen training distribution; they do not reduce to self-definition or imported uniqueness theorems. This matches the default non-circular case for standard ML benchmark papers.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard deep-learning training assumptions and the representativeness of the modified dataset; no new free parameters, axioms, or entities are introduced beyond those implicit in any CNN benchmark.

axioms (1)
  • domain assumption Cross-entropy loss and standard data-augmentation operations preserve semantic labels for galaxy morphology classes.
    Implicit in the training procedure described in the abstract.

pith-pipeline@v0.9.1-grok · 5719 in / 1351 out tokens · 32320 ms · 2026-06-27T18:38:03.264604+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

32 extracted references · 20 canonical work pages · 7 internal anchors

  1. [1]

    Millman, Rene , title =

  2. [2]

    Al-Sibai, Noor , title =

  3. [3]

    Surveying the Universe with Rubin Observatory

  4. [4]

    Astronomical Data Analysis Software and Systems XXVII , year = 2020, editor =

    Convolutional Neural Networks in Astronomy, and Applications for Diffuse Structure Discovery. Astronomical Data Analysis Software and Systems XXVII , year = 2020, editor =

    2020

  5. [5]

    No. 324. Extra-galactic nebulae. Contributions from the Mount Wilson Observatory / Carnegie Institution of Washington , year = 1926, month = jan, volume =

    1926

  6. [6]

    Leung, Henry , howpublished =

  7. [7]

    Overview of the DESI Legacy Imaging Surveys. Astron. J. , keywords =. doi:10.3847/1538-3881/ab089d , primaryClass =

    work page doi:10.3847/1538-3881/ab089d

  8. [8]

    Learning multiple layers of features from tiny images , year =

    Krizhevsky, Alex and Hinton, Geoffrey , address =. Learning multiple layers of features from tiny images , year =

  9. [9]

    The Sloan Digital Sky Survey: Technical Summary

    The Sloan Digital Sky Survey: Technical Summary. , keywords =. doi:10.1086/301513 , archivePrefix =. astro-ph/0006396 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1086/301513

  10. [10]

    Galaxy Zoo: Exploring the Motivations of Citizen Science Volunteers. Astron. Educ. Rev. , keywords =. doi:10.3847/AER2009036 , archivePrefix =. 0909.2925 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/aer2009036

  11. [11]

    , keywords =

    Galaxy Zoo DECaLS: Detailed visual morphology measurements from volunteers and deep learning for 314 000 galaxies. , keywords =. doi:10.1093/mnras/stab2093 , archivePrefix =. 2102.08414 , primaryClass =

    work page doi:10.1093/mnras/stab2093

  12. [12]

    and Vojtekova, Antonia and Anku, Anna and Walmsley, Mike and Garland, Izzy L

    O'Ryan, David and Merin, Bruno and Simmons, Brooke D. and Vojtekova, Antonia and Anku, Anna and Walmsley, Mike and Garland, Izzy L. and Geron, Tobias and Keel, William and Kruk, Sandor and Lintott, Chris J. and Mantha, Kameswara Bharadwaj and Masters, Karen L. and Reerink, Jan and Smethurst, Rebecca J. and Thorne, Matthew R. , title =. doi:10.5281/zenodo....

    work page doi:10.5281/zenodo.7684876

  13. [14]

    Monthly Notices of the Royal Astronomical Society , volume =

    Gharat, Sarvesh and Dandawate, Yogesh , title = ". Monthly Notices of the Royal Astronomical Society , volume =. 2022 , month =. doi:10.1093/mnras/stac457 , url =

    work page doi:10.1093/mnras/stac457 2022

  14. [15]

    Deep Galaxy: Classification of Galaxies based on Deep Convolutional Neural Networks

    Deep Galaxy: Classification of Galaxies based on Deep Convolutional Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.1709.02245 , archivePrefix =. 1709.02245 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1709.02245

  15. [16]

    and Al-Roainy, Ali A

    Ba Alawi, Abdulfattah E. and Al-Roainy, Ali A. , booktitle=. Deep Residual Networks Model for Star-Galaxy Classification , year=

  16. [17]

    Monthly Notices of the Royal Astronomical Society , volume =

    Cavanagh, Mitchell K and Bekki, Kenji and Groves, Brent A , title = ". Monthly Notices of the Royal Astronomical Society , volume =. 2021 , month =. doi:10.1093/mnras/stab1552 , url =

    work page doi:10.1093/mnras/stab1552 2021

  17. [18]

    , keywords =

    The Classification of Galaxy Morphology in the H Band of the COSMOS-DASH Field: A Combination-based Machine-learning Clustering Model. , keywords =. doi:10.3847/1538-4365/ace69e , archivePrefix =. 2307.02335 , primaryClass =

    work page doi:10.3847/1538-4365/ace69e

  18. [19]

    arXiv e-prints , keywords =

    Galaxy Morphology Classification using EfficientNet Architectures. arXiv e-prints , keywords =. doi:10.48550/arXiv.2008.13611 , archivePrefix =. 2008.13611 , primaryClass =

    work page doi:10.48550/arxiv.2008.13611 2008

  19. [20]

    2022 , month =

    Wuyu Hui and Zheng Robert Jia and Hansheng Li and Zijian Wang , title =. 2022 , month =. doi:10.1088/1742-6596/2402/1/012009 , url =

    work page doi:10.1088/1742-6596/2402/1/012009 2022

  20. [21]

    Galaxy morphology classification using VGG16

  21. [22]

    arXiv e-prints , keywords =

    Galaxy Classification Using Transfer Learning and Ensemble of CNNs With Multiple Colour Spaces. arXiv e-prints , keywords =. doi:10.48550/arXiv.2305.00002 , archivePrefix =. 2305.00002 , primaryClass =

    work page doi:10.48550/arxiv.2305.00002

  22. [23]

    Third International Conference on Computer Vision and Data Mining (ICCVDM 2022) , year = 2023, editor =

    Detection and classification of galaxy morphology based on YOLOv5. Third International Conference on Computer Vision and Data Mining (ICCVDM 2022) , year = 2023, editor =. doi:10.1117/12.2659975 , adsurl =

    work page doi:10.1117/12.2659975 2022

  23. [24]

    , keywords =

    YOLO-CL: Galaxy cluster detection in the SDSS with deep machine learning. , keywords =. doi:10.1051/0004-6361/202345976 , archivePrefix =. 2301.09657 , primaryClass =

    work page doi:10.1051/0004-6361/202345976

  24. [25]

    Rethinking the Inception Architecture for Computer Vision

  25. [26]

    Deep residual learning for image recognition , author=

  26. [27]

    Densely Connected Convolutional Networks

    Densely Connected Convolutional Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.1608.06993 , archivePrefix =. 1608.06993 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1608.06993

  27. [28]

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks

    EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks. arXiv e-prints , keywords =. doi:10.48550/arXiv.1905.11946 , archivePrefix =. 1905.11946 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1905.11946 1905

  28. [29]

    You Only Look Once: Unified, Real-Time Object Detection

    You Only Look Once: Unified, Real-Time Object Detection. arXiv e-prints , keywords =. doi:10.48550/arXiv.1506.02640 , archivePrefix =. 1506.02640 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1506.02640

  29. [30]

    Array programming with NumPy,

    Charles R. Harris and K. Jarrod Millman and St. Array programming with. 2020 , month = sep, journal =. doi:10.1038/s41586-020-2649-2 , publisher =

    work page doi:10.1038/s41586-020-2649-2 2020

  30. [31]

    10.5281/zenodo.3509134

    The pandas development team , title =. doi:10.5281/zenodo.3509134 , url =

    work page doi:10.5281/zenodo.3509134

  31. [32]

    Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning

    Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning. arXiv e-prints , keywords =. doi:10.48550/arXiv.1602.07261 , archivePrefix =. 1602.07261 , primaryClass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1602.07261

  32. [33]

    Al-Sibal, Noor , title =