Pith. sign in

REVIEW 3 major objections 6 minor 61 references

A dual-branch network that pairs a fixed scattering transform with learned convolutions, plus cropping by telescope beam size, best finds faint diffuse radio emission in galaxy clusters on small labelled samples.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-31 10:29 UTC pith:KDX65LRK

load-bearing objection Useful small-N recipe paper: DualSSN + beam crop + mild taper wins on LoTSS-DR2/PSZ2, but the ranking sits on a tiny fixed test set. the 3 major comments →

arxiv 2607.28349 v1 pith:KDX65LRK submitted 2026-07-30 astro-ph.CO

Multi-branch classification of diffuse cluster radio emission

Markus Bredberg , Emma Tolley This is my paper
classification astro-ph.CO
keywords galaxy clustersdiffuse radio emissionscattering transformsqueeze-excitationconvolutional neural networksLoTSSbeam-normalised croppingSKA
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Galaxy clusters sometimes glow with faint, megaparsec-scale radio emission that traces magnetic fields and merger-driven turbulence, but the signal sits only a few times above the noise and labelled examples are scarce. This paper asks how to classify cluster images as diffuse-emission or not when the training set is only a few hundred cutouts. It shows that a dual-branch classifier—one branch a standard convolutional network, the other operating on scattering-transform coefficients with squeeze-excitation attention—outperforms a plain CNN and a pure scattering network, and that the gain is largest when images are cropped to a fixed number of synthesised beams and mildly uv-tapered (or image-blurred) to enhance extended structure. Stacking several preprocessed versions of the same cluster does not help. The practical claim is that fixed multi-scale wavelet features plus beam-normalised cropping are a workable recipe for automated diffuse-emission finding ahead of SKA-scale surveys.

Core claim

On the small LoTSS-DR2/PSZ2 labelled sample, the DualSSN architecture (CNN branch plus scattering-transform branch with squeeze-excitation) combined with beam-normalised cropping and 25 kpc uv-tapering yields the strongest classification of diffuse versus non-diffuse cluster emission among the architectures and preprocessings tested, with mean accuracy about 0.86 over thirty runs and a top-five soft-vote ensemble at 0.94, while image-domain blurring performs comparably to true uv-tapering and multi-version stacking does not improve results.

What carries the argument

DualSSN: a dual-branch classifier that concatenates a learned convolutional encoder on the image with a squeeze-excitation convolutional encoder on fixed scattering-transform coefficients, so stable multi-scale wavelet invariants complement adaptive filters on limited data; paired with beam-normalised cropping that equalises the noise correlation scale across heterogeneous beams.

Load-bearing premise

That rankings measured on a fixed held-out test set of only a few dozen clusters, with high run-to-run variance and sometimes a single diffuse source in a validation fold, will still hold for SKA image products and other surveys.

What would settle it

Train and evaluate the same DualSSN versus CNN/ScatterNet baselines with beam versus FoV/pixel crops on an independent labelled cluster sample (for example another survey or a larger LoTSS release) using a freshly drawn test set; if beam-cropped DualSSN no longer leads on accuracy and F1, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Image-domain Gaussian blurring can substitute for uv-tapering when only restored maps are available, which matters for SKA-style archive products without routine visibility access.
  • Cropping classifiers to a fixed number of beams, not fixed pixels or fixed angular field of view, should be preferred when beam size varies across the sample.
  • Multi-branch designs that mix a fixed scattering front-end with a learned CNN branch are a concrete default for small, morphologically complex radio datasets rather than pure CNNs or pure scattering nets.
  • Stacking reference, blurred, and tapered versions of the same cluster as extra channels is not worth the added complexity on samples of this size.
  • Soft-voting a handful of high-accuracy runs can push single-run accuracy into the low-to-mid 0.9s on this task without changing the architecture.

Where Pith is reading between the lines

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

  • Beam-normalised cropping is effectively a physics-motivated domain adaptation step: it may transfer to other low-surface-brightness extended-source problems (e.g. residual Galactic emission or faint lobes) wherever the synthesised beam sets the noise texture.
  • The Grad-CAM split—CNN attending centre, scattering branch periphery—suggests a natural third branch (for example a vision transformer) could target intermediate scales the current pair under-use.
  • If labelled sets remain small, the same ST-plus-CNN pattern is a candidate for few-shot halo/relic/phoenix subclassification once those labels grow, before full segmentation models are needed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper benchmarks four classifiers (CNN, ScatterNet, DualCSN, DualSSN) for binary detection of diffuse cluster radio emission on the small LoTSS-DR2/PSZ2 labelled sample (207 clusters; 67 DE / 140 NDE). It combines dual-branch designs with scattering-transform features and squeeze-excitation attention, and systematically varies ten image versions (reference, blurred, uv-tapered, tapered+point-source-subtracted at 25/50/100 kpc) and three cropping strategies (pixel, fixed FoV, beam-normalised). The strongest reported configuration is DualSSN on beam-cropped 25 kpc uv-tapered images (mean accuracy 0.86±0.04 over 30 runs; top-5 soft-vote 0.94). The authors conclude that ST-based multi-branch models with beam-normalised cropping are a promising SKA-era direction, and that image-domain blurring is a viable substitute when visibilities are unavailable.

Significance. If the ranking holds under stronger validation, the work offers a practical, visibility-free preprocessing recipe (beam-normalised crop + mild taper/blur) and a compact dual-branch architecture well matched to the small-N regime of current diffuse-emission catalogues. Beam-normalised cropping is a clear, reusable contribution for radio classification. Code and LoTSS-DR2/PSZ2 data links are provided, supporting reproducibility. The claim is appropriately framed as a ‘promising direction’ rather than a definitive SKA pipeline, which fits the exploratory scope of A&A methods papers on survey readiness.

major comments (3)
  1. [§3.2, Table 4, Discussion] §3.2 and Table 4 (and Discussion, “Accuracy variation”): the headline ranking DualSSN + beam crop + Tap. 25 kpc (0.86±0.04; top-5 soft-vote 0.94) is measured on one fixed held-out test split of only ~13–17 DE and ~20–28 NDE clusters (Table 2), with validation folds that can contain a single DE source. The authors correctly note high run-to-run variance and that error bars may be underestimated because the test set is fixed. Soft-voting the top-5 runs ranked by that same test accuracy further selects on the evaluation set and inflates the 0.94 figure. With four architectures × many image versions × three crops, the winning cell is one of many correlated comparisons without nested/outer CV, test-set bootstrap, or multiplicity control. Please either (i) re-evaluate with nested CV or repeated random test splits and report the distribution of the DualSSN–baseline gap, or (ii) clearly demote t
  2. [§3.1, Figs. 4 and 6] §3.1–3.2 and Figs. 4, 6–7: RadioUNet (Stuardi et al. 2024) is plotted as a single black point and repeatedly called a ‘rough contextual benchmark,’ yet the text still states that DualSSN ‘outperforms RadioUNet for all three cropping settings.’ The comparison mixes different inputs (~6″ Botteon products vs 20″ archive), tasks (image-level label vs segmentation threshold), and training regimes (real-only vs synthetic pretrain + fine-tune). Either remove the numerical outperformance language and keep RadioUNet strictly as context, or add a matched evaluation (same images, same split, image-level labels only) so the gap is interpretable.
  3. [§2.4.3, Appendix B, Fig. B.1] §2.4.3 and Appendix B: beam-normalised cropping is motivated by standardising the noise correlation scale, which is a strong and novel idea, but the implementation equalises only the global mean FoV across versions and still leaves per-cluster and per-version FoV variation (and non-circular beams). Figure B.1 shows systematic differences between tapered and blurred target beams. Please quantify how much residual beam-count or noise-texture variation remains after cropping (e.g. distribution of beams per image and of effective noise correlation length), and test whether the DualSSN advantage over DualCSN/CNN shrinks when all models see strictly matched beam counts. Without that check, the causal link ‘beam crop → uniform noise scale → better morphology discrimination’ remains partly conjectural.
minor comments (6)
  1. [Table 2, §2.4.1] Table 2: training/validation ranges (e.g. DE train 46–51) are clear, but state explicitly whether the fixed test set is stratified by RH vs RR and by redshift, and whether the 12 RH+RR systems are always counted once as DE.
  2. [§2.5] §2.5: label-smoothing ε is introduced symbolically but its numerical value is not given alongside mixup α=β=0.4, lr, and λ_L2. Please list ε in the training hyperparameter summary.
  3. [Fig. 8, §3.2] Fig. 8 Grad-CAM panels: the probability annotations and run indices are useful; add a short note on whether Grad-CAM was computed on the merged logit or per branch before fusion, since the figure overlays both branches.
  4. [Table C.1, Appendix C] Table C.1 pools many configurations and is valuable; clarify in the caption that pixel-crop rows are excluded and that means are unweighted over cells (so versions with more trained crops do not dominate).
  5. [Abstract, §2.1] Typos/notation: abstract ‘uv- tapering’ stray space; §2.1 ‘T ransform’; consistent use of Tap. vs Tap and Blur vs Blurred in figure legends.
  6. [§4] §4 Generalisability: the decision not to retune hyperparameters per configuration is defensible; still, a one-sentence sensitivity check (e.g. ±factor of two in lr or L2 on the winning cell) would strengthen the claim that DualSSN’s edge is architectural rather than tuning luck.

Circularity Check

0 steps flagged

No circularity: empirical supervised ranking on external Botteon labels; architecture/preprocessing comparisons do not reduce predictions to fitted inputs by construction.

full rationale

This paper is a standard small-data ML benchmark. Labels (RH/RR/NDE vs NDE) are taken from Botteon et al. (2022) on LoTSS-DR2/PSZ2; models (CNN, ScatterNet, DualCSN, DualSSN) and preprocessing (beam/FoV/pixel crop; reference/blur/tap/sub at 25–100 kpc) are compared via accuracy, precision, recall, F1, and AUC on a fixed held-out test split with cross-validation folds. Nothing in the derivation chain defines a quantity in terms of the quantity it claims to predict, fits a parameter and renames the fit a prediction, or imports a uniqueness theorem that forces the DualSSN+beam+25 kpc ranking. The mild self-reference to Tolley (2024) only motivates ScatterNet’s fixed ST+MLP design and parameter count; it does not underwrite the empirical ranking or forbid alternatives. Soft-voting top-N runs ranked on the same test set can inflate a reported number, and the tiny fixed test set raises robustness concerns, but those are statistical validity issues, not circular reduction of claim to input. The central conclusion—that ST multi-branch models with beam-normalised cropping are a promising SKA-era direction—is an empirical preference statement, not a first-principles derivation. Score 0; steps empty.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 2 invented entities

Load-bearing content is experimental ML practice plus standard radio-interferometric preprocessing, not new physical postulates. Free parameters are training/preprocessing knobs held mostly fixed across configs. Axioms are domain assumptions about labels, beams, and blur–taper equivalence. No new physical entities.

free parameters (5)
  • percentile clip bounds (p_lo=0.3, p_hi=0.99) and arcsinh α=10 = 0.3 / 0.99 / α=10
    Per-image normalisation hyperparameters chosen for training; authors note global-range normalisation did not help.
  • learning rate, L2, label-smoothing ε, mixup α=β, dropout schedule, batch size 16 = lr=4e-5, λ_L2=0.1, mixup α=β=0.4
    Optimiser and regularisation settings fixed rather than retuned per configuration (lr=4e-5, λ_L2=0.1, mixup 50% with β(0.4,0.4)).
  • ST parameters J=2, L=12, order 2; network widths/depths = J=2, L=12, m≤2
    Scattering and CNN capacity choices that set feature dimensionality and parameter counts (e.g. DualSSN 116k params).
  • beam-count equalisation target and FoV θ=800 arcsec / pixel 512 crop = θ=800''; pixel 512; n*_V globally equalised
    Cropping hyperparameters that define the three competing input geometries; beam multipliers calibrated globally to match mean angular FoV.
  • taper/blur physical scales {25,50,100} kpc = 25, 50, 100 kpc
    Discrete resolution scales taken from Botteon products / matched blur; performance depends on this choice (25 kpc best).
axioms (5)
  • domain assumption Botteon et al. (2022) definite RH/RR/NDE labels (candidates excluded) are a sufficiently reliable binary ground truth for supervised training.
    Entire supervised benchmark rests on these 207 labelled clusters (67 DE / 140 NDE).
  • domain assumption Cropping to a fixed number of synthesised-beam FWHMs approximately equalises noise correlation scale across heterogeneous LoTSS beams and tapers.
    Stated motivation for beam cropping (Methods 2.4.3); residual non-circular beams and NaN caps are acknowledged in App. B.
  • domain assumption Under imperfect uv-coverage, image-domain Gaussian convolution is still a useful proxy for visibility-plane Gaussian taper for classification (not identical imaging).
    Appendix B derives ideal equivalence then notes cleaning-order and coverage differences; results treat blur as viable SKA-era substitute.
  • domain assumption Downsampling cropped cutouts to 128×128 preserves large-scale diffuse structure relevant to the DE/NDE decision.
    Supported by authors’ azimuthally averaged power-spectrum check above the Nyquist scale of the resized maps.
  • standard math Standard supervised learning assumptions: cross-entropy with label smoothing/mixup, AdamW, and train/val/test splits yield meaningful generalisation estimates on this sample.
    Ordinary ML training theory; no exotic loss derivation.
invented entities (2)
  • DualSSN (dual-branch scatter squeeze network) no independent evidence
    purpose: Named architecture combining CNN image branch with ST+SE branch for DE/NDE classification.
    Engineering composition of known ST, SE, and dual-branch ideas; not a new physical object. Independent evidence is the empirical accuracy tables only.
  • Beam-normalised cropping strategy for radio classification no independent evidence
    purpose: Fix input geometry in units of local effective beam to standardise noise texture.
    Authors claim novelty in radio-astronomy classification context; it is a preprocessing rule, not a physical entity. Evidence is internal ablation only.

pith-pipeline@v1.2.0-daily-grok45 · 33560 in / 4048 out tokens · 66133 ms · 2026-07-31T10:29:37.399035+00:00 · methodology

0 comments
read the original abstract

Context. Galaxy clusters sometimes host synchrotron radiation on scales of ~100 kpc to ~1 Mpc, with surface brightness only a few times the image noise. This diffuse cluster radio emission is a sensitive probe of magnetic fields and intracluster medium dynamics, but disentangling the underlying physical processes requires statistically large samples spanning a wide range of cluster masses, dynamical states, and redshifts, together with sufficient sensitivity to low-surface-brightness emission. Aims. We explore two techniques for improving detection of diffuse emission in galaxy cluster images, relative to a baseline classifier: the scattering transform (ST) and squeeze-excitation (SE) attention. Methods. We integrate an ST encoder into a dual-branch classifier (DualSSN) and a scattering network (ScatterNet). We incorporate SE attention into the DualSSN and dual-branch convolutional neural network (DualCSN). These classifiers are then benchmarked against a simple CNN, across ten image preprocessing configurations and three cropping strategies. Performance is evaluated on small labelled datasets from the second data release of the LOFAR two-metre sky survey overlapping with the second Planck catalogue of Sunyaev-Zel'dovich sources (LoTSS-DR2/PSZ2). Results. Alongside the multi-branch approach with SE and ST, cropping the image to a fixed number of telescope beams and uv- tapering (smoothing to a coarser angular resolution) improve classification performance, while stacking multiple preprocessed ver- sions of an image does not. Conclusions. Scattering-transform-based multi-branch architectures with beam-normalised cropping are a promising direction for diffuse emission classification in the SKA era.

Figures

Figures reproduced from arXiv: 2607.28349 by Emma Tolley, Markus Bredberg.

Figure 1
Figure 1. Figure 1: Schematic view of the four different classifier models used in this work. MLP: multi-layered perceptron. SE: Squeeze￾excitation. eralise well across the full range of source morphologies. In both architectures, one branch is identical to the simple CNN, and the other branches apply an SE block after each convolutional block. The two architectures differ in their input to the second branch. The third model,… view at source ↗
Figure 2
Figure 2. Figure 2: Different steps of the data preprocessing. three point-source-subtracted versions, per image. The point￾source-subtracted images of type TapXkpc will be referred to as Tap.+Sub. X kpc. We use all three LoTSS-DR2 image types in this work because tapered image files show larger-scale emission more clearly and are expected to improve classification performance. Every image cutout is labelled as either diffuse… view at source ↗
Figure 3
Figure 3. Figure 3: Examples of images processed under three cropping strategies (rows) across seven image versions (columns), for one ran [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: , and also includes the accuracy of Stuardi et al. (2024) shown as a single black dot. Their result of 0.73 was obtained on the same LoTSS-DR2/PSZ2 sample but used the standard 20′′ low-resolution archive images, whereas the reference images used here are the post-processed ∼6 ′′ images from Botteon et al. (2022). The two performances are therefore not directly compa￾rable, and the RadioUNet accuracy is in… view at source ↗
Figure 5
Figure 5. Figure 5: shows the corresponding confusion matrices. The confusion matrix of DualSSN shows a more balanced error dis￾tribution between false positives and false negatives than the other models. 3.2. Blurred, tapered and point-source-subtracted data [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Classification accuracy for all four classifiers and three [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: DualSSN classification accuracy across tapering/ [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Six correctly classified images (left) and six erroneously classified images (right) from the DualSSN on 25 kpc [PITH_FULL_IMAGE:figures/full_fig_p009_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Confusion matrices comparing image versions using DualSSN on beam-cropped data (four matrices to the left), cropping [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: ROC curves comparing (left) dataset versions using DualSSN on beam-cropped data, (middle) cropping modes for 25 kpc [PITH_FULL_IMAGE:figures/full_fig_p010_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Size and noise stability of the four classifier architec [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗
Figure 3
Figure 3. Figure 3: We see two major differences between the uv-tapered images and our blurred images. First, the blurred images are noisier than the uv-tapered im￾ages. This is likely partly a consequence of the different sky re￾gions in the versions combined with a per-image normalisation (see Section 2.4.4). If the uv-tapered images contain a larger por￾tion of high-intensity pixels, slightly brighter low-intensity pix￾els… view at source ↗

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

61 extracted references · 6 linked inside Pith

  1. [1]

    2019, A&A, 629, A115

    Allys, E., Levrier, F., Zhang, S., et al. 2019, A&A, 629, A115

    2019

  2. [2]

    2020, Phys

    Allys, E., Marchand, T., Cardoso, J.-F., et al. 2020, Phys. Rev. D, 102, 103506

    2020

  3. [3]

    Aniyan, A. K. & Thorat, K. 2017, The Astrophysical Journal Supplement Series, 230, 20

    2017

  4. [4]

    2025, SKAO Science Data Products: A Summary, Technical Report SKA-TEL-SKO-0001818, SKA Observatory, Jodrell Bank, UK, revision 02

    Arumugam, V ., Breen, S., Bolton, R., et al. 2025, SKAO Science Data Products: A Summary, Technical Report SKA-TEL-SKO-0001818, SKA Observatory, Jodrell Bank, UK, revision 02

    2025

  5. [5]

    & Hwang, H

    Bahk, H. & Hwang, H. S. 2024, ApJS, 272, 7

    2024

  6. [6]

    W., Cassano, R., et al

    Botteon, A., Shimwell, T. W., Cassano, R., et al. 2022, A&A, 660, A78

    2022

  7. [7]

    Briggs, D. S. 1995, in American Astronomical Society Meeting Abstracts, V ol. 187, American Astronomical Society Meeting Abstracts, 112.02 Brüggen, M., Bykov, A., Ryu, D., & Röttgering, H. 2012, Space Sci. Rev., 166, 187

    1995

  8. [8]

    & Mallat, S

    Bruna, J. & Mallat, S. 2012, arXiv e-prints, arXiv:1203.1513

    Pith/arXiv arXiv 2012

  9. [9]

    A., Grandón, D., et al

    Cheng, S., Marques, G. A., Grandón, D., et al. 2024, arXiv e-prints, arXiv:2404.16085

    Pith/arXiv arXiv 2024

  10. [10]

    & Ménard, B

    Cheng, S. & Ménard, B. 2021, arXiv e-prints, arXiv:2112.01288

    Pith/arXiv arXiv 2021

  11. [11]

    2023, arXiv e-prints, arXiv:2306.17210

    Cheng, S., Morel, R., Allys, E., Ménard, B., & Mallat, S. 2023, arXiv e-prints, arXiv:2306.17210

    Pith/arXiv arXiv 2023

  12. [12]

    2020, Monthly Notices of the Royal Astronomical Society, 499, 5902

    Cheng, S., Ting, Y .-S., Ménard, B., & Bruna, J. 2020, Monthly Notices of the Royal Astronomical Society, 499, 5902

    2020

  13. [13]

    E., Kronberg, P

    Clarke, T. E., Kronberg, P. P., & Böhringer, H. 2001, ApJ, 547, L111

    2001

  14. [14]

    2012, A&A Rev., 20, 54 García-Farieta, J

    Feretti, L., Giovannini, G., Govoni, F., & Murgia, M. 2012, A&A Rev., 20, 54 García-Farieta, J. E., Hortúa, H. J., & Kitaura, F.-S. 2024, A&A, 684, A100

    2012

  15. [15]

    2018, MNRAS, 480, 3749

    Gheller, C., Vazza, F., & Bonafede, A. 2018, MNRAS, 480, 3749

    2018

  16. [16]

    2020, A&A, 640, A108

    Giovannini, G., Cau, M., Bonafede, A., et al. 2020, A&A, 640, A108

    2020

  17. [17]

    2016, Deep Learning (MIT Press), http://www.deeplearningbook.org

    Goodfellow, I., Bengio, Y ., & Courville, A. 2016, Deep Learning (MIT Press), http://www.deeplearningbook.org

    2016

  18. [18]

    & Feretti, L

    Govoni, F. & Feretti, L. 2004, International Journal of Modern Physics D, 13, 1549

    2004

  19. [19]

    2009, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edn

    Hastie, T., Tibshirani, R., & Friedman, J. 2009, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd edn. (Springer)

    2009

  20. [20]

    N., Brüggen, M., Botteon, A., et al

    Hoang, D. N., Brüggen, M., Botteon, A., et al. 2022, A&A, 665, A60

    2022

  21. [21]

    2018, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Hu, J., Shen, L., & Sun, G. 2018, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2018

  22. [22]

    2015, in Advancing Astro- physics with the Square Kilometre Array (AASKA14), 92

    Johnston-Hollitt, M., Govoni, F., Beck, R., et al. 2015, in Advancing Astro- physics with the Square Kilometre Array (AASKA14), 92

    2015

  23. [23]

    S., Mudigere, D., Nocedal, J., Smelyanskiy, M., & Tang, P

    Keskar, N. S., Mudigere, D., Nocedal, J., Smelyanskiy, M., & Tang, P. T. P. 2017, in International Conference on Learning Representations

    2017

  24. [24]

    2021, in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, 1098–1106

    Kinakh, V ., Taran, O., & V oloshynovskiy, S. 2021, in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Work- shops, 1098–1106

    2021

  25. [25]

    D., Rudnick, L., et al

    Knowles, K., Cotton, W. D., Rudnick, L., et al. 2022, A&A, 657, A56

    2022

  26. [26]

    2025, MNRAS, 543, 1638

    Kolokythas, K., Venturi, T., Knowles, K., et al. 2025, MNRAS, 543, 1638

    2025

  27. [27]

    Krizhevsky, A., Sutskever, I., & Hinton, G. E. 2012, in Advances in Neural Infor- mation Processing Systems, ed. F. Pereira, C. Burges, L. Bottou, & K. Wein- berger, V ol. 25 (Curran Associates, Inc.)

    2012

  28. [28]

    & Hertz, J

    Krogh, A. & Hertz, J. 1991, in Advances in Neural Information Processing Sys- tems, ed. J. Moody, S. Hanson, & R. Lippmann, V ol. 4 (Morgan-Kaufmann)

    1991

  29. [29]

    G., Waterson, M., Alachkar, B., et al

    Labate, M. G., Waterson, M., Alachkar, B., et al. 2022, Journal of Astronomical

    2022

  30. [30]

    2017, in Advances in Neu- ral Information Processing Systems, ed

    Lakshminarayanan, B., Pritzel, A., & Blundell, C. 2017, in Advances in Neu- ral Information Processing Systems, ed. I. Guyon, U. V . Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, & R. Garnett, V ol. 30 (Curran As- sociates, Inc.)

    2017

  31. [31]

    & Clark, S

    Lei, M. & Clark, S. E. 2023, The Astrophysical Journal, 947, 74

    2023

  32. [32]

    2025, arXiv e-prints, arXiv:2508.19311

    Li, D., Liu, T., Liu, A., et al. 2025, arXiv e-prints, arXiv:2508.19311

    arXiv 2025

  33. [33]

    & Hutter, F

    Loshchilov, I. & Hutter, F. 2019, in International Conference on Learning Rep- resentations

    2019

  34. [34]

    1999, in Proceedings of the Seventh IEEE International Conference on Computer Vision, V ol

    Lowe, D. 1999, in Proceedings of the Seventh IEEE International Conference on Computer Vision, V ol. 2, 1150–1157 vol.2

    1999

  35. [35]

    2012, Communications on Pure and Applied Mathematics, 65, 1331

    Mallat, S. 2012, Communications on Pure and Applied Mathematics, 65, 1331

    2012

  36. [36]

    2024, Astronomy and Computing, 48, 100835 Maslej-Krešˇnáková, V ., El-Bouchefry, K., & Butka, P

    Manzano, C., Miskolczi, A., Stiele, H., et al. 2024, Astronomy and Computing, 48, 100835 Maslej-Krešˇnáková, V ., El-Bouchefry, K., & Butka, P. 2021, Monthly Notices of the Royal Astronomical Society, 505, 1464

    2024

  37. [37]

    K., Tolley, E., Krishna, S

    Mishra, A. K., Tolley, E., Krishna, S. P., & Kneib, J.-P. 2025, MNRAS, 538, 2905

    2025

  38. [38]

    2022, arXiv e-prints, arXiv:2204.10177

    Morel, R., Rochette, G., Leonarduzzi, R., Bouchaud, J.-P., & Mallat, S. 2022, arXiv e-prints, arXiv:2204.10177

    Pith/arXiv arXiv 2022

  39. [39]

    A., et al

    Mousset, L., Allys, E., Price, M. A., et al. 2024, arXiv e-prints, arXiv:2407.07007 Müller, R., Kornblith, S., & Hinton, G. E. 2019, in Advances in Neural Infor- mation Processing Systems, ed. H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, & R. Garnett, V ol. 32 (Curran Associates, Inc.)

    Pith/arXiv arXiv 2024

  40. [40]

    & Hinton, G

    Nair, V . & Hinton, G. E. 2010, in Proceedings of the 27th International Confer- ence on Machine Learning (Omnipress), 807–814 Ndung’u, S., Grobler, T., Wijnholds, S. J., Karastoyanova, D., & Azzopardi, G. 2023, New Astronomy Reviews, 97, 101685

    2010

  41. [41]

    Norris, R. P. 2016, Proceedings of the International Astronomical Union, 12, 103–113

    2016

  42. [42]

    L., et al

    Oktay, O., Schlemper, J., Folgoc, L. L., et al. 2018, in Medical Imaging with Deep Learning

    2018

  43. [43]

    J., Boxelaar, J

    Osinga, E., van Weeren, R. J., Boxelaar, J. M., et al. 2021, A&A, 648, A11

    2021

  44. [44]

    Pacholczyk, A. G. 1970, Radio astrophysics. Nonthermal processes in galactic and extragalactic sources (W. H. Freeman) Planck Collaboration, Ade, P. A. R., Aghanim, N., et al. 2016, A&A, 594, A27 pytorch developers. 2025,pytorch.nn.functional.interpolate(Docu- mentation),https://docs.pytorch.org/docs/stable/generated/ torch.nn.functional.interpolate.html,...

    1970

  45. [45]

    2025, MNRAS, 541, 3479

    Sanvitale, N., Gheller, C., Vazza, F., et al. 2025, MNRAS, 541, 3479

    2025

  46. [46]

    K., Portillo, S

    Saydjari, A. K., Portillo, S. K. N., Slepian, Z., et al. 2021, ApJ, 910, 122

    2021

  47. [47]

    R., Cogswell, M., Das, A., et al

    Selvaraju, R. R., Cogswell, M., Das, A., et al. 2017, in Proceedings of the IEEE International Conference on Computer Vision (ICCV)

    2017

  48. [48]

    W., Hardcastle, M

    Shimwell, T. W., Hardcastle, M. J., Tasse, C., et al. 2022, A&A, 659, A1

    2022

  49. [49]

    W., Röttgering, H

    Shimwell, T. W., Röttgering, H. J. A., Best, P. N., et al. 2017, A&A, 598, A104

    2017

  50. [50]

    2014, Journal of Machine Learning Research, 15, 1929

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. 2014, Journal of Machine Learning Research, 15, 1929

    2014

  51. [51]

    2021, Monthly Notices of the Royal Astronomical Society, 502, 2518

    Stuardi, C., Bonafede, A., Lovisari, L., et al. 2021, Monthly Notices of the Royal Astronomical Society, 502, 2518

    2021

  52. [52]

    2025, A&A, 695, L16

    Stuardi, C., Botteon, A., Sereno, M., et al. 2025, A&A, 695, L16

    2025

  53. [53]

    2024, MNRAS, 533, 3194

    Stuardi, C., Gheller, C., Vazza, F., & Botteon, A. 2024, MNRAS, 533, 3194

    2024

  54. [54]

    P., Dewdney, P

    Swart, G. P., Dewdney, P. E., & Cremonini, A. 2022, Journal of Astronomical

    2022

  55. [55]

    2016, in Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Szegedy, C., Vanhoucke, V ., Ioffe, S., Shlens, J., & Wojna, Z. 2016, in Proceed- ings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2016

  56. [56]

    2025, A&A, 701, A114

    Tevlin, L., Berlok, T., Pfrommer, C., et al. 2025, A&A, 701, A114

    2025

  57. [57]

    2024, URSI Radio Science Letters, 5, 28 van Haarlem, M

    Tolley, E. 2024, URSI Radio Science Letters, 5, 28 van Haarlem, M. P., Wise, M. W., Gunst, A. W., et al. 2013, A&A, 556, A2 van Weeren, R. J., de Gasperin, F., Akamatsu, H., et al. 2019, Space Sci. Rev., 215, 16

    2024

  58. [58]

    & SKA Science Working Groups

    Wagg, J. & SKA Science Working Groups. 2021, SKA1 Scientific Use Cases, Technical Report SKA-TEL-SKO-0000015, SKA Observatory, revision 04

    2021

  59. [59]

    2023, Universe, 9, 319

    Wittor, D. 2023, Universe, 9, 319

    2023

  60. [60]

    Woo, S., Park, J., Lee, J.-Y ., & Kweon, I. S. 2018, in Proceedings of the European Conference on Computer Vision (ECCV)

    2018

  61. [61]

    N., & Lopez-Paz, D

    Zhang, H., Cisse, M., Dauphin, Y . N., & Lopez-Paz, D. 2018, in International Conference on Learning Representations Article number, page 12 Bredberg & Tolley: Multi-branch classification of diffuse cluster radio emission Table A.1: CNN architecture. Layer Component Depth Act. Reg. (p) Feature Extractor 1–2 3×3 Conv, BN 8 LReLU DO2d (0.3) 3–4 3×3 Conv, BN...

    2018