pith. sign in

arxiv: 2606.11595 · v1 · pith:2IMV2ATAnew · submitted 2026-06-10 · 🌌 astro-ph.IM · gr-qc

Wavelet-Based Extraction of Transient Noise in Gravitational-Wave Interferometers using a Saliency-Guided Learning Architecture

Christopher Allene , Dhruv Kumar , Yusuke Sakai , Marco Meyer-Conde , Hirotaka Takahashi This is my paper

Pith reviewed 2026-06-27 08:39 UTC · model grok-4.3

classification 🌌 astro-ph.IM gr-qc
keywords gravitational wave interferometerstransient noise extractionglitchescontinuous wavelet transformdiscrete wavelet transformsaliency mapsmachine learningwaveform reconstruction
0
0 comments X

The pith

A saliency-guided transfer from continuous to discrete wavelet transforms extracts and reconstructs transient glitches in gravitational-wave data.

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

The paper presents a supervised method that first tags glitch candidates with UMAP, then uses a model on continuous wavelet transform spectrograms to generate saliency maps marking the relevant time-frequency regions. Those maps are mapped onto an invertible discrete wavelet transform basis so that coefficient masking can produce both the isolated glitch waveform and the cleaned strain data with no reconstruction error. The approach targets families such as whistles and scattered-light transients and is shown to handle low signal-to-noise cases and partial overlaps where simple thresholding or band-pass filters leak energy or miss structure. Because the final step is linear and invertible, the extracted signals remain usable for downstream detector characterization or subtraction tasks. A sympathetic reader would see this as supplying an interpretable, scalable route from raw strain to cleanly separated glitch and astrophysical content.

Core claim

The central claim is that saliency patterns learned on continuous wavelet transform spectrograms can be transferred to an invertible discrete wavelet transform representation, where adaptive coefficient masking yields exact reconstruction of both glitch-only and glitch-suppressed waveforms across multiple glitch families and in regimes where classical methods fail.

What carries the argument

Saliency-guided transfer between continuous and discrete wavelet transforms followed by adaptive coefficient masking in the discrete basis.

If this is right

  • Exact glitch subtraction becomes possible without residual leakage into the cleaned strain.
  • The same saliency maps can serve as diagnostics for learned representations via UMAP.
  • The method scales to larger glitch catalogs because it avoids non-invertible operations after the saliency step.
  • Low signal-to-noise and partially overlapping transients can be handled without the failures typical of thresholding or band-limited filters.

Where Pith is reading between the lines

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

  • The framework could be extended to real-time subtraction pipelines if the discrete-wavelet masking step is made sufficiently fast.
  • Because the reconstruction is exact, the extracted glitches could be used directly as training targets for future morphology classifiers without additional denoising.
  • The approach may generalize to other non-stationary noise sources in interferometric data if the saliency model is retrained on new families.

Load-bearing premise

Saliency patterns identified on continuous wavelet transform spectrograms transfer to the discrete wavelet transform without introducing significant artifacts or information loss that would prevent exact reconstruction.

What would settle it

Apply the pipeline to a set of low signal-to-noise overlapping whistle and scattered-light glitches and measure whether the reconstructed glitch-only waveform matches the injected transient to within the noise floor after subtraction.

Figures

Figures reproduced from arXiv: 2606.11595 by Christopher Allene, Dhruv Kumar, Hirotaka Takahashi, Marco Meyer-Conde, Yusuke Sakai.

Figure 1
Figure 1. Figure 1: FIG. 1: Workflow overview for transient-noise analysis using preprocessed gravitational-wave strain data. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: Schematic representation of the data-augmentation strategy applied during training. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: Example of Scattered Light glitch ( [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4 [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: FIG. 5: This example illustrates the ability of the method to transfer time-frequency saliency into an invertible multi [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: FIG. 6: UMAP latent-space representations in two dimensions based on the learned feature embeddings. The dot [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: FIG. 7 [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
read the original abstract

Gravitational-wave interferometers exhibit a wide variety of short-duration non-Gaussian transients, commonly referred to as glitches, that complicate the detection of astrophysical signals, bias parameter estimation, and detector characterisation. Existing machine-learning approaches classify glitch morphologies but do not provide a complete mechanism to segment and extract these disturbances from the strain data. We introduce a wavelet-based, saliency-guided framework for the supervised extraction of transient noise. Candidates are first pre-tagged using Uniform Manifold Approximation and Projection, which is also used as a diagnostic of the learned representations. A traditional learning model operating on Continuous Wavelet Transform spectrograms then identifies relevant time-frequency regions through saliency maps. These saliency patterns are transferred to an invertible multiresolution representation via the Discrete Wavelet Transform, where adaptive coefficient masking enables exact reconstruction of both glitch-only and glitch-suppressed waveforms. We demonstrate effective extraction across several representative glitch families, including 'Whistle' and 'Scattered-Light' transients, and show robustness in challenging regimes such as low signal-to-noise events and partially overlapping structures, where classical thresholding or band-limited filtering methods typically fail or introduce leakage. The proposed framework offers an interpretable and computationally efficient approach to transient-noise extraction, establishing a foundation for scalable applications to larger glitch catalogs and future observing runs.

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 / 0 minor

Summary. The paper proposes a wavelet-based saliency-guided supervised framework for extracting transient glitches from gravitational-wave strain data. Candidates are pre-tagged with UMAP; saliency maps on CWT spectrograms identify time-frequency regions; these patterns are transferred to an invertible DWT representation for adaptive coefficient masking, enabling claimed exact reconstruction of glitch-only and glitch-suppressed waveforms. The method is demonstrated on representative families such as 'Whistle' and 'Scattered-Light' transients and asserted to be robust in low-SNR and overlapping regimes where classical methods fail.

Significance. If the central extraction claims hold after validation, the framework would supply an interpretable, computationally efficient alternative to existing glitch-classification ML methods, with potential utility for improving strain data quality ahead of future observing runs. The use of UMAP as both pre-tagger and diagnostic is a positive design choice.

major comments (2)
  1. [Abstract] Abstract (pipeline description): The claim of 'exact reconstruction of both glitch-only and glitch-suppressed waveforms' rests on transfer of saliency patterns from CWT spectrograms to DWT coefficient masks, yet no transfer operator, interpolation rule, or invertibility proof is supplied. Because CWT is redundant while DWT is critically sampled, any non-bijective mapping risks support misalignment and leakage into the cleaned strain; this step is load-bearing for the extraction guarantee but remains undescribed.
  2. [Abstract] Abstract: The assertions of 'effective extraction across several representative glitch families' and 'robustness in challenging regimes such as low signal-to-noise events and partially overlapping structures' are presented without any quantitative metrics, error bars, baseline comparisons (e.g., against wavelet thresholding or existing ML denoisers), or cross-validation details. The central claim of superiority therefore cannot be assessed from the supplied information.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment point by point below and will revise the manuscript to improve clarity and support for the central claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract (pipeline description): The claim of 'exact reconstruction of both glitch-only and glitch-suppressed waveforms' rests on transfer of saliency patterns from CWT spectrograms to DWT coefficient masks, yet no transfer operator, interpolation rule, or invertibility proof is supplied. Because CWT is redundant while DWT is critically sampled, any non-bijective mapping risks support misalignment and leakage into the cleaned strain; this step is load-bearing for the extraction guarantee but remains undescribed.

    Authors: We agree that the abstract does not describe the transfer operator, interpolation rule, or invertibility argument. The manuscript body outlines the overall pipeline but does not provide the explicit mapping details or proof. We will add a dedicated subsection to the Methods section that formalizes the saliency-to-DWT mask transfer (including the mathematical operator and any interpolation), along with a discussion of exact reconstruction under adaptive masking and the conditions under which support misalignment is avoided. revision: yes

  2. Referee: [Abstract] Abstract: The assertions of 'effective extraction across several representative glitch families' and 'robustness in challenging regimes such as low signal-to-noise events and partially overlapping structures' are presented without any quantitative metrics, error bars, baseline comparisons (e.g., against wavelet thresholding or existing ML denoisers), or cross-validation details. The central claim of superiority therefore cannot be assessed from the supplied information.

    Authors: The abstract is a high-level summary and does not include quantitative metrics. The manuscript demonstrates the method on Whistle and Scattered-Light families with qualitative examples but lacks the requested quantitative metrics, error bars, baselines, and cross-validation details in the results. We will revise the abstract to include key quantitative results and expand the Results section with explicit metrics (e.g., suppression ratios, reconstruction errors), baseline comparisons to wavelet thresholding, error bars, and cross-validation details to support the claims. revision: yes

Circularity Check

0 steps flagged

No significant circularity; pipeline is self-contained methodological proposal

full rationale

The paper introduces a new supervised extraction framework combining UMAP pre-tagging, saliency mapping on CWT spectrograms, and transfer to invertible DWT for coefficient masking. No equations, fitted parameters, or self-citations are presented that reduce any claimed performance or reconstruction guarantee to a self-referential definition or input. The derivation chain consists of independent algorithmic steps whose validity rests on empirical demonstration rather than tautology. The transfer operator between CWT saliency and DWT masks is described as part of the method but is not shown to be equivalent to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review limits visibility into parameters and assumptions. No explicit free parameters, new entities, or non-standard axioms are stated; the pipeline relies on standard wavelet and ML components whose details are not provided.

pith-pipeline@v0.9.1-grok · 5780 in / 1226 out tokens · 15600 ms · 2026-06-27T08:39:24.666015+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

49 extracted references · 1 canonical work pages

  1. [1]

    Q-Transform representation The primary representation used by Omicron is the Q-transform, evaluated around trigger times and imple- mented inGWpy. For a strain time seriesx(t), the Q- 3 Dataset Golden Dataset Preprocess Pretagging (a) Training - Clipping augmentation - Dilation augmentation - Validation & Test Preprocess (b) Classification -Inference, CAM...

  2. [2]

    The CWT of a signalx(t) is defined as Wx(a, b) = 1 |a| 1 p Z +∞ −∞ x(t)ψ ∗ t−b a dt,(2) whereaandbdenote the scale and time shift, respec- tively, andψis the chosen mother wavelet

    Continuous Wavelet Transform representation Wavelet-based representations provide a complemen- tary description of transient structures through localized decompositions across multiple scales. The CWT of a signalx(t) is defined as Wx(a, b) = 1 |a| 1 p Z +∞ −∞ x(t)ψ ∗ t−b a dt,(2) whereaandbdenote the scale and time shift, respec- tively, andψis the chosen...

  3. [3]

    The model used in this study is based on aResNet-50 architecture[30] withImageNet-pretrained weights[31], imposing a fixed input size of 224×224 pixels

    Learning architecture and Pre-tagging. The model used in this study is based on aResNet-50 architecture[30] withImageNet-pretrained weights[31], imposing a fixed input size of 224×224 pixels. The network begins with an initial convolutional block con- sisting of a 7×7 convolution, batch normalisation, ReLU activation, and a 3×3 max-pooling layer. It is fo...

    2048

  4. [4]

    During training, we vary the clipping threshold of the spectrogram amplitudes by a value randomly drawn from a uniform distribution over the interval [3,23]

    Training The training dataset consists of a balanced sample of 4,500 Gravity Spy glitches drawn from three rep- resentative morphologies:Whistle,Blip, andScattered LightThe dataset is split into training, validation, and test subsets with proportions of 70%, 20%, and 10%, respectively. During training, we vary the clipping threshold of the spectrogram amp...

  5. [5]

    The way the data augmentation apply during the inference phase is describe on Fig

    Inference for glitch denoising Two distinct inference procedures are employed. The way the data augmentation apply during the inference phase is describe on Fig. 2b. In the first stage, the ob- jective is to localize the glitch within the input. A single clipping threshold, fixed at 13, is used while all four pre- defined durations (0.5, 1, 2, and 4 s) ar...

  6. [6]

    CAM selection and combination Beyond classification, this architecture is ex- tended to localize transient structures by coupling GradCAM++[22] to generate two-dimensional saliency maps that highlight the regions contributing most strongly to the classification decision. The Grad- CAM++ using higher order derivative may fail to compute a map, as a fallbac...

  7. [7]

    After the flattening, to select the crop, we search for the excesses of signal above the average value

    CAM-based glitch localization The CAM obtained is flatten along frequency axis tak- ing the maximum values to estimate an effective glitch duration, which is used to define a refined temporal crop ∆tof the spectrogram. After the flattening, to select the crop, we search for the excesses of signal above the average value. Ranking the excesses by there maxi...

  8. [8]

    Mask construction The cCAM is used to reweight the input spectrogram in order to enhance regions relevant to the glitch while suppressing unrelated high-amplitude features. The weighted spectrogram is defined as: wpx,ij = q inputij ×CAM ij.(7) wherew px,ij denotes the weighted pixel value at posi- tion (i, j), input ij is the original spectrogram intensit...

    arXiv

  9. [9]

    The red dashed box is the coarse saliency bounding box of width ∆tand height ∆f.(c)The computation of the mask

    The mother wavelet used is a complex Morlet with the parameter 0.25 and 8 with a L2-norm.(b)The CAM corresponding to the inference on the input image. The red dashed box is the coarse saliency bounding box of width ∆tand height ∆f.(c)The computation of the mask. The two plots included show an example of distributions of CAM pixel’s inside (red) or outside...

    2048

  10. [10]

    T. L. S. Collaboration, Classical and Quantum Gravity 32, 074001 (2015)

    2015

  11. [11]

    Acernese and other, Classical and Quantum Gravity 32, 024001 (2014)

    F. Acernese and other, Classical and Quantum Gravity 32, 024001 (2014)

    2014

  12. [12]

    KAGRA Collaboration, Progress of Theoretical and Ex- perimental Physics2021, 05A101 (2020)

    2020

  13. [13]

    B. P. Abbott and other (LIGO Scientific Collaboration and Virgo Collaboration), Phys. Rev. Lett.119, 161101 (2017)

    2017

  14. [14]

    Bahaadini, V

    S. Bahaadini, V. Noroozi, N. Rohani, S. Coughlin, M. Zevin, J. Smith, V. Kalogera, and A. Katsaggelos, Information Sciences444, 172 (2018)

    2018

  15. [15]

    Sakai, Y

    Y. Sakai, Y. Itoh, P. Jung, K. Kokeyama, C. Kozakai, K. T. Nakahira, S. Oshino, Y. Shikano, H. Takahashi, 14 T. Uchiyama, G. Ueshima, T. Washimi, T. Yamamoto, and T. Yokozawa, Scientific Reports12, 9935 (2022)

    2022

  16. [16]

    Zevin, S

    M. Zevin, S. Coughlin, S. Bahaadini, E. Besler, N. Ro- hani, S. Allen, M. Cabero, K. Crowston, A. K. Katsagge- los, S. L. Larson, T. K. Lee, C. Lintott, T. B. Littenberg, A. Lundgren, C. Østerlund, J. R. Smith, L. Trouille, and V. Kalogera, Classical and Quantum Gravity34, 064003 (2017)

    2017

  17. [17]

    McInnes, J

    L. McInnes, J. Healy, and J. Melville, Umap: Uniform manifold approximation and projection for dimension re- duction (2020), arXiv:1802.03426 [stat.ML]

    Pith/arXiv arXiv 2020

  18. [18]

    Oshino, Y

    S. Oshino, Y. Sakai, M. Meyer-Conde, T. Uchiyama, Y. Itoh, Y. Shikano, Y. Terada, and H. Takahashi, Physics Letters B870, 139938 (2025)

    2025

  19. [19]

    Robinet, N

    F. Robinet, N. Arnaud, N. Leroy, A. Lundgren, D. Macleod, and J. McIver, SoftwareX12, 100620 (2020)

    2020

  20. [20]

    Daubechies, IEEE Transactions on Information Theory 36, 961 (1990)

    I. Daubechies, IEEE Transactions on Information Theory 36, 961 (1990)

    1990

  21. [21]

    Virtuoso and E

    A. Virtuoso and E. Milotti, Phys. Rev. D109, 102010 (2024)

    2024

  22. [22]

    Pati and P

    Y. Pati and P. Krishnaprasad, inAdvances in Neural In- formation Processing Systems, Vol. 3, edited by R. Lipp- mann, J. Moody, and D. Touretzky (Morgan-Kaufmann, 1990)

    1990

  23. [23]

    N. J. Cornish, Phys. Rev. D102, 124038 (2020)

    2020

  24. [24]

    N. J. Cornish, T. B. Littenberg, B. B´ ecsy, K. Chatziioan- nou, J. A. Clark, S. Ghonge, and M. Millhouse, Physical Review D103, 10.1103/physrevd.103.044006 (2021)

    work page doi:10.1103/physrevd.103.044006 2021

  25. [25]

    Parida and N

    P. Parida and N. Bhoi, Future Computing and Informat- ics Journal2, 65 (2017)

    2017

  26. [26]

    B. P. Abbott and other (LIGO Scientific Collaboration and Virgo Collaboration), Phys. Rev. X9, 031040 (2019)

    2019

  27. [27]

    Abbott and other (LIGO Scientific Collaboration and Virgo Collaboration), Phys

    R. Abbott and other (LIGO Scientific Collaboration and Virgo Collaboration), Phys. Rev. X11, 021053 (2021)

    2021

  28. [28]

    T. L. S. Collaboration and the Virgo Collaboration, Gwtc-2.1: Deep extended catalog of compact binary coa- lescences observed by ligo and virgo during the first half of the third observing run (2022), arXiv:2108.01045 [gr- qc]

    Pith/arXiv arXiv 2022

  29. [29]

    Abbott and other (LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration), Phys

    R. Abbott and other (LIGO Scientific Collaboration, Virgo Collaboration, and KAGRA Collaboration), Phys. Rev. X13, 041039 (2023)

    2023

  30. [30]

    B. Zhou, A. Khosla, A. Lapedriza, A. Oliva, and A. Tor- ralba, Learning deep features for discriminative localiza- tion (2015), arXiv:1512.04150 [cs.CV]

    Pith/arXiv arXiv 2015

  31. [31]

    Chattopadhay, A

    A. Chattopadhay, A. Sarkar, P. Howlader, and V. N. Bal- asubramanian, in2018 IEEE Winter Conference on Ap- plications of Computer Vision (WACV)(2018) pp. 839– 847

    2018

  32. [32]

    KAGRA Collaboration, Progress of Theoretical and Ex- perimental Physics2023, 10A101 (2022)

    2022

  33. [33]

    Redmon, S

    J. Redmon, S. K. Divvala, R. B. Girshick, and A. Farhadi, CoRRabs/1506.02640(2015), 1506.02640

    Pith/arXiv arXiv 2015

  34. [34]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin, At- tention is all you need (2023), arXiv:1706.03762 [cs.CL]

    Pith/arXiv arXiv 2023

  35. [35]

    Zhang, X

    Z. Zhang, X. Lu, G. Cao, Y. Yang, L. Jiao, and F. Liu, in 2021 IEEE/CVF International Conference on Computer Vision Workshops (ICCVW)(2021) pp. 2799–2808

    2021

  36. [36]

    Lilly and S

    J. Lilly and S. Olhede, IEEE Transactions on Signal Pro- cessing57, 146–160 (2009)

    2009

  37. [37]

    J. M. Lilly, Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences473, 20160776 (2017)

    2017

  38. [38]

    Kingsbury, Applied and Computational Harmonic Analysis10, 234 (2001)

    N. Kingsbury, Applied and Computational Harmonic Analysis10, 234 (2001)

    2001

  39. [39]

    K. He, X. Zhang, S. Ren, and J. Sun, Deep residual learning for image recognition (2015), arXiv:1512.03385 [cs.CV]

    Pith/arXiv arXiv 2015

  40. [40]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei- Fei, in2009 IEEE Conference on Computer Vision and Pattern Recognition(2009) pp. 248–255

    2009

  41. [41]

    R. Fu, Q. Hu, X. Dong, Y. Guo, Y. Gao, and B. Li, Axiom-based grad-cam: Towards accurate visualiza- tion and explanation of cnns (2020), arXiv:2008.02312 [cs.CV]

    arXiv 2020

  42. [42]

    S. G. Chang, C. SG, B. Yu, Y. B, M. Vetterli, and V. M, IEEE transactions on image processing : a publication of the IEEE Signal Processing Society9, 1532 (2000)

    2000

  43. [43]

    D. L. Donoho and I. M. Johnstone, Biometrika81, 425 (1994)

    1994

  44. [44]

    Bajaj, inWavelet Theory, edited by S

    N. Bajaj, inWavelet Theory, edited by S. Mohammady (IntechOpen, London, 2020) Chap. 5

    2020

  45. [45]

    Mazumder, M

    M. Mazumder, M. Assaduzzaman, M. Islam, and M. S. Hossain, Current Journal of Applied Science and Tech- nology28, 1 (2018)

    2018

  46. [46]

    K M and K

    J. K M and K. S., Journal of Applied Research and Tech- nology16, 103 (2018)

    2018

  47. [47]

    Donoho, IEEE Transactions on Information Theory 41, 613 (1995)

    D. Donoho, IEEE Transactions on Information Theory 41, 613 (1995)

    1995

  48. [48]

    Koyama, Y

    N. Koyama, Y. Sakai, S. Sasaoka, D. Dominguez, K. Somiya, Y. Omae, Y. Terada, M. Meyer-Conde, and H. Takahashi, Machine Learning: Science and Technol- ogy5, 035028 (2024)

    2024

  49. [49]

    S. Soni, N. Mukund, and E. Katsavounidis, Gw-yolo: Multi-transient segmentation in ligo using computer vi- sion (2025), arXiv:2508.17399 [astro-ph.IM]

    arXiv 2025