Pith. sign in

REVIEW 4 major objections 5 minor 82 references

ArtifactGen: Benchmarking WGAN-GP vs Diffusion for Label-Aware EEG Artifact Synthesis

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read For 1–2 second EEG artifact windows, a conditional WGAN-GP matches real spectra better than a diffusion baseline, though both models remain detectable.

desk verdict Picks a genuinely useful benchmark question and ships a real pipeline, but the central WGAN-vs-diffusion ranking is asserted rather than shown—no numbers anywhere, and the authors concede the comparison is confounded. read the letter →

arxiv 2509.08188 v1 pith:FDMKR53V submitted 2025-09-09 cs.LG cs.NEq-bio.NC

Hritik Arasu , Faisal R Jahangiri This is my paper
classification cs.LGcs.NEq-bio.NC
keywords EEGartifactsynthesisconditionalWGAN-GPdiffusionmodelsclassifier-freeguidancespectralfidelitymaximummeandiscrepancydataaugmentationtime-seriesgeneration
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper asks whether modern generative models can produce realistic, label-aware EEG artifact segments that could serve as training data. Using an annotated EEG artifact corpus, it splits subjects into training, validation, and test sets, and compares a conditional WGAN-GP with a projection discriminator against a 1D denoising diffusion model with classifier-free guidance. Its central claim is that, under its configuration, the WGAN-GP achieves closer spectral alignment and lower distributional distance (MMD) to real data than the diffusion model, while both models show weak class-conditioned recovery and leave detectable traces in temporal and cross-channel structure. If that holds, it means the adversarial baseline is currently the stronger choice for short-window artifact synthesis, but neither approach is yet reliable enough for augmentation, and the released pipeline gives the community a controlled setup to test improvements.

What carries the argument

The central mechanism is the head-to-head comparison itself: a conditional WGAN-GP with a projection discriminator (which scores real versus fake windows by an inner product between learned features and class embeddings) against a 1D denoising diffusion model with FiLM-style label conditioning and classifier-free guidance. The evaluation machinery consists of time-series-aware fidelity metrics—Welch band-power deltas, channel-covariance Frobenius distance, autocorrelation L2 distance, and MMD—computed on subject-disjoint 8-channel windows at 250 Hz.

What would settle it

Run the released pipeline with the diffusion model matched to the WGAN's 1-second windows and per-window min–max normalization, using several hundred sampling steps and a tuned higher-order sampler; if diffusion then achieves equal or better band-power deltas and MMD, the paper's ranking would be refuted.

Watch

Extended reading notes

Core claim

On five artifact classes—muscle, eye, electrode, chewing, and shiver—the paper reports that the WGAN-GP's synthetic windows have smaller relative error in band power across δ/θ/α/β/γ and a lower maximum mean discrepancy to real windows than the diffusion model's. It attributes the advantage to per-window min–max normalization and shorter 1-second windows that act as a spectral regularizer, plus a projection discriminator that injects class labels directly into the critic. The diffusion model used z-scored normalization, longer windows, a limited number of sampling steps, and a modest 1D U-Net, and the authors state this comparison is not perfectly controlled. Both models nevertheless produce

Load-bearing premise

The WGAN's reported advantage over diffusion could be an artifact of unequal configurations—different window lengths, normalization, sampling steps, and model capacities—so the load-bearing premise is that the comparison is fair enough to attribute the gap to the generative approach.

Editorial extensions

If this is right

  • If the WGAN-GP result is correct, a projection-conditioned adversarial model is a stronger baseline than a modestly configured diffusion model for short-window EEG artifact synthesis, so future comparisons should beat the WGAN-GP, not just the diffusion baseline.
  • Both models leave detectable distribution shift: a simple nearest-neighbor classifier in feature space can separate synthetic from real windows, so synthetic artifacts should not yet be used for augmentation without post-hoc filtering.
  • Weak class-conditional recovery means current models do not reliably preserve artifact identity, limiting their use for label-aware stress-testing.
  • Domain-specific metrics (band-power deltas, covariance/ACF distances, MMD) ranked models consistently, while image-style metrics like PRD were unstable, suggesting EEG evaluation should emphasize spectral and structural measures.
  • For diffusion to close the gap, the paper recommends more sampling steps or higher-order samplers, EDM-style parameterization, careful guidance scaling, and montage-aware architectures.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: The ranking could reverse if the diffusion model were given the same window length, normalization, and more sampling steps; the paper's own limitations section says the comparison is not perfectly controlled, so the honest reading is that the WGAN's advantage is configuration-dependent, not a fundamental property of adversarial versus diffusion training.
  • Editorial inference: A practical test of this would be rerunning the released pipeline with a matched 1-second, min–max-scaled diffusion configuration; if diffusion then matches or beats the WGAN on band-power deltas and MMD, the paper's headline ranking would be refuted.
  • Editorial inference: The weak class-conditional recovery suggests that neither model has learned the mapping from label to artifact morphology well enough; future gains may come less from architecture and more from explicit spectral/topographic losses and stronger conditioning, not from scaling either baseline.
  • Editorial inference: The finding that even the 'winning' synthetic samples are detectable by a simple classifier implies that any downstream augmentation use should include a realism filter or acceptance test, and the released metrics could be used as a regression test for future generators.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. ArtifactGen proposes a benchmark for label-aware EEG artifact synthesis, comparing a conditional WGAN-GP with a projection discriminator against a 1D DDPM with classifier-free guidance on TUH EEG Artifact (TUAR) data. The paper describes subject-wise data splits, model-specific preprocessing (per-window min-max for WGAN, per-recording z-score for diffusion), a set of signal-level evaluation metrics (Welch band-power errors, covariance/ACF distances, MMD, diversity), and a reproducibility pipeline with manifests and configs. The central claim, stated in the abstract and Section 6, is that in this setting WGAN-GP achieves closer spectral alignment and lower MMD to real data than the diffusion baseline, while both models show weak class-conditional recovery and leave detectable temporal/cross-channel traces.

Significance. If fully supported, a reproducible head-to-head comparison of conditional GAN and diffusion models for EEG artifact synthesis would be a useful community baseline, particularly given the paper's emphasis on subject-wise splits, domain-appropriate metrics, and released code/configs. However, the significance cannot currently be assessed because the paper's central empirical ranking is presented only qualitatively: no tables, numeric results, error bars, or statistical tests appear for the claimed MMD, band-power errors, covariance distances, or class-recovery numbers. In addition, the authors acknowledge major confounds (window length, normalization, sampling steps, capacity) that are entangled with model identity, so even the qualitative ranking cannot be attributed to WGAN-GP versus diffusion as model classes. The reproducibility artifacts are a strength, but they do not compensate for the absence of the benchmark's actual measurements.

major comments (4)
  1. [§5.4 and §6] The evaluation section defines RelErr, PSD L2, MMD, diversity, and channel-mean discrepancies, and states 'These statistics are the basis of our quantitative comparisons.' Yet Section 6 reports only qualitative assertions—'consistently lower relative band-power errors,' 'a smaller MMD to the real distribution'—with no numerical values, tables, figures, or appendix entries. The reader cannot verify the direction, magnitude, or statistical significance of the WGAN-vs-diffusion differences claimed in the abstract. This is a load-bearing omission for a benchmark paper.
  2. [§6 Limitations] The limitations paragraph explicitly acknowledges that window length (250 vs 500), normalization (per-window min-max vs per-recording z-score), sampling steps (50), guidance tuning, and U-Net capacity differ between models. These factors vary jointly with model class, so any observed spectral/MMD ordering cannot be cleanly attributed to WGAN-GP versus diffusion. The paper's central claim is therefore not supported by a controlled comparison, even if the missing numbers were supplied.
  3. [§6 vs §A.1] There is a direct inconsistency in the reported diffusion sampling configuration. Section 6 states the diffusion configuration used '50 sampling steps and v-prediction,' while Appendix A.1 states 'sampling with 80-step deterministic DDIM-style schedule' and describes no v-prediction. This ambiguity undermines reproducibility and prevents the reader from knowing which configuration produced the claimed results.
  4. [Abstract and §1 vs §5.4/§6] The abstract and introduction promise evaluation along a utility axis ('augmentation effects on artifact recognition') and feature-space metrics (FID/KID/PRD). However, Section 5.4 defines only signal-level and distributional metrics plus a diversity proxy, and Section 6 does not report any augmentation or downstream artifact-recognition results; FID/KID/PRD values are also absent. Several promised evaluation components are missing from the reported study.
minor comments (5)
  1. [§1] Duplicate citation '[22, 22]' in the introduction; should be a single reference.
  2. [§5.4] The diversity score '1 − corr' depends on the representation and window length; the text should specify the feature representation used and whether the score is bounded or normalized.
  3. [§6] The term 'v-prediction' is introduced without definition, and Section 5.2 describes the diffusion model as predicting additive noise (epsilon-prediction). Clarify the training objective.
  4. [§5.4] Per-channel mean discrepancies are said to be provided for channels c=0,...,4, but the montage has eight channels; clarify whether only five channels are reported and why.
  5. [General] Reference formatting is inconsistent (mixed capitalization, some entries missing venue/arXiv IDs); a final copyediting pass is needed for the journal style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning identified: the paper's comparative claims are empirical, evaluated against external real-data metrics; no load-bearing self-citation or definitional reduction is present.

full rationale

The paper does not contain a mathematical derivation chain that could reduce to its inputs. Its central claim—that WGAN-GP achieves closer spectral alignment and lower MMD than diffusion—is an empirical comparison evaluated with external metrics (Welch band-power, MMD, covariance/ACF distances) computed against real EEG data. These metrics are not defined in terms of the models' own outputs or fitted parameters. The only hyperparameter tuning mentioned (CFG scale tuned on validation FID, Appendix A.1) is a standard model-selection step and does not constitute a circular prediction of the reported spectral/MMD comparisons. The Limitations section explicitly acknowledges confounds (window length, normalization, sampling steps, model capacity), which is an experimental-fairness concern, not a circularity concern; it does not disguise the result as a derivation. There are no self-citations that bear the load of the argument, no uniqueness theorems imported from the authors' prior work, and no ansatz smuggled in via citation that predetermines the outcome. The absence of a numeric results table is a reporting/verifiability weakness but does not make the claim circular. Overall, the central result is externally anchored and self-contained against real-data benchmarks, so the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central comparison rests on a set of modeling and preprocessing choices that are not derived from theory. The paper itself flags several as limitations (window lengths, normalization, sampling steps, U-Net capacity). No new physical or conceptual entities are introduced.

free parameters (6)
  • Diffusion window length = 500 samples (2 s)
    The diffusion model uses L=500 while WGAN uses L=250; this difference is a confound in the comparison.
  • Diffusion CFG scale = 1.5
    Tuned on validation FID (Appendix A.1), affects the spectral alignment and diversity of diffusion samples.
  • Diffusion sampling steps = 50 or 80 (inconsistent)
    Section 6 says 50, Appendix A.1 says 80; this directly affects sample quality and is a confound.
  • Diffusion U-Net capacity = channel widths (64,128,256)
    Modest capacity, the authors list it as a limitation affecting the diffusion comparison.
  • WGAN window length = 250 samples (1 s)
    Adversarial path uses 1-second windows; the authors hypothesize this favors WGAN.
  • Normalization schemes = per-window min-max (WGAN) vs per-recording/channel z-score (diffusion)
    Different preprocessing per model, acknowledged as a confound in Section 6.
assumptions (4)
  • domain assumption The 8-channel montage Fp1, Fp2, C3, C4, O1, O2, T3, T4 captures sufficient artifact morphology.
    Section 4: all recordings are reduced to these channels; artifacts with other spatial patterns are outside scope.
  • domain assumption Welch band-power deltas and MMD are sufficient fidelity metrics for the comparison.
    Section 5.4 and 6: the authors themselves note temporal and cross-channel structure is not captured, so this axiom is partially acknowledged.
  • domain assumption Raw signals without notch/band-pass filtering preserve artifact morphology.
    Section 4: 'we operate on raw signals to preserve artifact morphology'.
  • domain assumption Subject-wise splits prevent leakage.
    Section 4: they enforce 149/32/32 subject splits; assumes annotations and recordings are correctly associated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArtifactGen: Benchmarking WGAN-GP vs Diffusion for Label-Aware EEG Artifact Synthesis." pith.science (2026). https://pith.science/paper/FDMKR53V

@misc{pith2026250908188,
  author       = {Pith},
  title        = {Pith review of: ArtifactGen: Benchmarking WGAN-GP vs Diffusion for Label-Aware EEG Artifact Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FDMKR53V}},
  note         = {Machine review of arXiv:2509.08188}
}
abstract

Artifacts in electroencephalography (EEG) -- muscle, eye movement, electrode, chewing, and shiver -- confound automated analysis yet are costly to label at scale. We study whether modern generative models can synthesize realistic, label-aware artifact segments suitable for augmentation and stress-testing. Using the TUH EEG Artifact (TUAR) corpus, we curate subject-wise splits and fixed-length multi-channel windows (e.g., 250 samples) with preprocessing tailored to each model (per-window min-max for adversarial training; per-recording/channel $z$-score for diffusion). We compare a conditional WGAN-GP with a projection discriminator to a 1D denoising diffusion model with classifier-free guidance, and evaluate along three axes: (i) fidelity via Welch band-power deltas ($\Delta\delta,\ \Delta\theta,\ \Delta\alpha,\ \Delta\beta$), channel-covariance Frobenius distance, autocorrelation $L_2$, and distributional metrics (MMD/PRD); (ii) specificity via class-conditional recovery with lightweight $k$NN/classifiers; and (iii) utility via augmentation effects on artifact recognition. In our setting, WGAN-GP achieves closer spectral alignment and lower MMD to real data, while both models exhibit weak class-conditional recovery, limiting immediate augmentation gains and revealing opportunities for stronger conditioning and coverage. We release a reproducible pipeline -- data manifests, training configurations, and evaluation scripts -- to establish a baseline for EEG artifact synthesis and to surface actionable failure modes for future work.

Figures

Figures reproduced from arXiv: 2509.08188 by the authors.

Figure 1
Figure 1. Representative multi-channel EEG windows for each artifact class, illustrating morphology [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Distributional alignment in embedding space. Comparison of (a) t [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Additional qualitative example of the shiver class. Multi-channel windows highlighting [PITH_FULL_IMAGE:figures/full_fig_p015_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Per-file embedding summaries. t-SNE (a) and UMAP (b) projections aggregated per recording, illustrating within-file cluster structure and variability [PITH_FULL_IMAGE:figures/full_fig_p016_4.png]
Figure 5
Figure 5. Figure 5: Channel distribution per split (multilabel). Relative presence of channels across train/val/test, [PITH_FULL_IMAGE:figures/full_fig_p016_5.png]
Figure 6
Figure 6. Figure 6: Window duration statistics by artifact (multilabel). Boxplots summarize duration dispersion, [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 57 canonical work pages

  1. [1]

    Data augmentation of eeg using gans for emotion recognition.IEEE Transactions on Affective Computing, 2021

    Hamed Azami, Mostafa Rostaghi, and Javier Escudero. Data augmentation of eeg using gans for emotion recognition.IEEE Transactions on Affective Computing, 2021

    2021

  2. [2]

    Engemann, and Alexandre Gramfort

    Hubert Banville, Omar Chehab, Aapo Hyvärinen, Denis A. Engemann, and Alexandre Gramfort. uncovering the structure of clinical eeg signals with self-supervised learning.Nature Machine Intelligence, 3:873–881, 2021

    2021

  3. [3]

    classification of covariance matrices using riemannian geometry and its application to brain-computer interfaces

    Alexandre Barachant, Stephane Bonnet, Marco Congedo, and Christian Jutten. classification of covariance matrices using riemannian geometry and its application to brain-computer interfaces. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 21(4):532–543, 2013

    work page 2013

  4. [4]

    Nima Bigdely-Shamlo, Tim Mullen, Christian Kothe, Kyung-Mi Su, and Kay A. Robbins. The prep pipeline: Standardized preprocessing for large-scale eeg analysis.Frontiers in Neuroinformatics, 9:16, 2015

    work page 2015

  5. [5]

    Demystifying MMD GANs

    Mikolaj Binkowski, Dougal J Sutherland, Michael Arbel, and Arthur Gretton. Demystifying MMD GANs. InInternational Conference on Learning Representations, 2018

    work page 2018

  6. [6]

    John Buckwalter et al. Recent advances in the tuh eeg corpus: Improving the interrater agreement for artifacts and epileptiform events.2021 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), pages 1–5, 2021

    work page 2021

  7. [7]

    The secret sharer: Evaluating and testing unintended memorization in neural networks

    Nicholas Carlini, Chang Liu, Úlfar Erlingsson, Jernej Kos, and Dawn Song. The secret sharer: Evaluating and testing unintended memorization in neural networks. In28th USENIX Security Symposium (USENIX Security 19), pages 267–284, Santa Clara, CA, aug 2019. USENIX Association

    work page 2019

  8. [8]

    Extracting training data from diffusion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramèr, Borja Balle, Daphne Ippolito, and Eric Wallace. Extracting training data from diffusion models. In32nd USENIX Security Symposium (USENIX Security 23), pages 5253–5270, Anaheim, CA, aug 2023. USENIX Association

    work page 2023

Show all 82 references
  1. [9]

    Improving eeg classification through randomly reassembling original and generated data with transformer- based diffusion models.arXiv preprint arXiv:2407.20253, 2024

    Mingzhi Chen, Yiyu Gui, Yuqi Su, Yuesheng Zhu, Guibo Luo, and Yuchao Yang. Improving eeg classification through randomly reassembling original and generated data with transformer- based diffusion models.arXiv preprint arXiv:2407.20253, 2024

    2024 arXiv

  2. [10]

    Weiss, Mohammad Norouzi, and William Chan

    Nanxin Chen, Yu Zhang, Heiga Zen, Ron J. Weiss, Mohammad Norouzi, and William Chan. wavegrad: estimating gradients for waveform generation.arxiv preprint arxiv:2009.00713, 2020

    2009 arXiv

  3. [11]

    Eeg artifact simulation for training robust deep networks.Frontiers in Neuroinformatics, 16:835699, 2022

    Pascal Croce, Bin Yang, Tilmann Sander, and Nicolas Langer. Eeg artifact simulation for training robust deep networks.Frontiers in Neuroinformatics, 16:835699, 2022

    2022

  4. [12]

    A comparative analysis in eeg artifact detection.arXiv preprint arXiv:2401.05409, 2023

    Binayak Das, Vaibhav Gandhi, et al. A comparative analysis in eeg artifact detection.arXiv preprint arXiv:2401.05409, 2023

    2023 arXiv

  5. [13]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. InAdvances in Neural Information Processing Systems (NeurIPS), 2021

    2021

  6. [14]

    differentially private diffusion models.arxiv preprint arxiv:2210.12101, 2022

    Tim Dockhorn, Arash Vahdat, and Karsten Kreis. differentially private diffusion models.arxiv preprint arxiv:2210.12101, 2022

    2022 arXiv

  7. [15]

    Adversarial audio synthesis

    Chris Donahue, Julian McAuley, and Miller Puckette. Adversarial audio synthesis. InInterna- tional Conference on Learning Representations, 2019

    2019

  8. [16]

    Jinhao Duan, Fei Kong, Shiqi Wang, Xiaoshuang Shi, and Kaidi Xu. Are diffusion models vulnerable to membership inference attacks? InProceedings of the 40th International Confer- ence on Machine Learning, volume 202 ofProceedings of Machine Learning Research, pages 8717–8730. P...

    2023

  9. [17]

    Hyland, and Gunnar Rätsch

    Cristóbal Esteban, Stephanie L. Hyland, and Gunnar Rätsch. Real-valued (medical) time series generation with recurrent conditional gans. InAdvances in Neural Information Processing Systems (NeurIPS), 2017

    2017

  10. [18]

    Goldberger, Luis A

    Ary L. Goldberger, Luis A. N. Amaral, Leon Glass, Jeffrey M. Hausdorff, Plamen Ch. Ivanov, Roger G. Mark, Joseph E. Mietus, George B. Moody, Chung-Kang Peng, and H. Eugene Stanley. physiobank, physiotoolkit, and physionet: components of a new research resource for complex phys...

    2000

  11. [19]

    a kernel two-sample test.Journal of Machine Learning Research, 13(25):723–773, 2012

    Arthur Gretton, Karsten Borgwardt, Malte Rasch, Bernhard Schölkopf, and Alexander Smola. a kernel two-sample test.Journal of Machine Learning Research, 13(25):723–773, 2012

    2012

  12. [20]

    Improved training of wasserstein gans

    Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron C Courville. Improved training of wasserstein gans. InAdvances in Neural Information Processing Systems, 2017

    2017

  13. [21]

    Habashi, Ahmed M

    Ahmed G. Habashi, Ahmed M. Azab, Seif Eldawlatly, and Gamal M. Aly. Generative adversarial networks in EEG analysis: an overview.Journal of NeuroEngineering and Rehabilitation, 20, apr 2023

    2023

  14. [22]

    Hamid, K

    A. Hamid, K. Gagliano, S. Rahman, N. Tulin, V . Tchiong, I. Obeid, and J. Picone. The temple university artifact corpus: An annotated corpus of eeg artifacts. In2020 IEEE Signal Processing in Medicine and Biology Symposium (SPMB), pages 1–4, 2020

    2020

  15. [23]

    Eeg-gan: Genera- tive adversarial networks for electroencephalographic (eeg) brain signals.arXiv preprint arXiv:1806.01875, 2018

    Kay Gregor Hartmann, Robin Tibor Schirrmeister, and Tonio Ball. Eeg-gan: Genera- tive adversarial networks for electroencephalographic (eeg) brain signals.arXiv preprint arXiv:1806.01875, 2018

    2018 arXiv

  16. [24]

    deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. deep residual learning for image recognition. Inproceedings of the ieee conference on computer vision and pattern recognition (cvpr), pages 770–778, 2016

    2016

  17. [25]

    Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshmi- narayanan

    Dan Hendrycks, Norman Mu, Ekin D. Cubuk, Barret Zoph, Justin Gilmer, and Balaji Lakshmi- narayanan. Augmix: A simple data processing method to improve robustness and uncertainty. InInternational Conference on Learning Representations, 2020

    2020

  18. [26]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. InAdvances in Neural Information Processing Systems, 2017

    2017

  19. [27]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, 2020

    2020

  20. [28]

    Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance.arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  21. [29]

    Energy-efficient tree-based eeg artifact detection.arXiv preprint arXiv:2204.09577, 2022

    Thorir Már Ingolfsson, Andrea Cossettini, Simone Benatti, and Luca Benini. Energy-efficient tree-based eeg artifact detection.arXiv preprint arXiv:2204.09577, 2022

    2022 arXiv

  22. [30]

    Jackson and Luis M

    Christopher W. Jackson and Luis M. Bolanos. eeg artifact removal and synthesis for clinical use.Frontiers in Neuroscience, 15:674681, 2021

    2021

  23. [31]

    Removal of artifacts from eeg signals: A review

    Xinyang Jiang, Guobao Bian, and Zhen Tian. Removal of artifacts from eeg signals: A review. Sensors, 19(5):987, 2019

    2019

  24. [32]

    Aneeg: Leveraging deep learning for effective artifact removal in eeg data.Scientific Reports, 14(1):14761, 2024

    Zubayer Kalita, Apoorv Kumar, Jagadeesh Yadav, and Deepshikha. Aneeg: Leveraging deep learning for effective artifact removal in eeg data.Scientific Reports, 14(1):14761, 2024

    2024

  25. [33]

    elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Samuli Laine, Erik Härkönen, Joni-Petteri Hellsten, Jaakko Lehtinen, and Timo Aila. elucidating the design space of diffusion-based generative models. Inadvances in neural information processing systems (neurips), 2022

    2022

  26. [34]

    The tuh abnormal expansion eeg corpus (tuabex).NeuroImage: Clinical, 38:103317, 2023

    Julia Kiessner et al. The tuh abnormal expansion eeg corpus (tuabex).NeuroImage: Clinical, 38:103317, 2023. 10

    2023

  27. [35]

    frechet audio distance: a metric for evaluating music enhancement algorithms.arxiv preprint arxiv:1812.08466, 2019

    Kevin Kilgour, Mauricio Zuluaga, Umut Simsekli, Justin Salamon, Juan Pablo Bello, and Najim Dehak. frechet audio distance: a metric for evaluating music enhancement algorithms.arxiv preprint arxiv:1812.08466, 2019

    2019 arXiv

  28. [36]

    Adam: A method for stochastic optimization

    Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInterna- tional Conference on Learning Representations, 2015

    2015

  29. [37]

    Synthesizing eeg signals from event-related potential paradigms with conditional diffusion models

    Guido Klein, Pierre Guetschel, Gianluigi Silvestri, and Michael Tangermann. Synthesizing eeg signals from event-related potential paradigms with conditional diffusion models. In Proceedings of the 9th Graz Brain–Computer Interface Conference, pages 438–443, 2024

    2024

  30. [38]

    diffwave: a versatile diffusion model for audio synthesis.arxiv preprint arxiv:2009.09761, 2020

    Zhifeng Kong, Wei Ping, Jiaji Huang, Kexin Zhao, and Bryan Catanzaro. diffwave: a versatile diffusion model for audio synthesis.arxiv preprint arxiv:2009.09761, 2020

    2009 arXiv

  31. [39]

    improved precision and recall metric for assessing generative models

    Tuomas Kynkäänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. improved precision and recall metric for assessing generative models. Inadvances in neural information processing systems (neurips), 2019

    2019

  32. [40]

    Lawhern, Amelia J

    Vernon J. Lawhern, Amelia J. Solon, Nicholas R. Waytowich, Shaun M. Gordon, Chou P. Hung, and Brent J. Lance. eegnet: a compact convolutional neural network for eeg-based brain–computer interfaces.Journal of Neural Engineering, 15(5):056013, 2018

    2018

  33. [41]

    Diffusion models for time series applications: A survey.arXiv preprint arXiv:2305.00624, 2023

    Lequan Lin, Zhengkun Li, Ruikun Li, Xuliang Li, and Junbin Gao. Diffusion models for time series applications: A survey.arXiv preprint arXiv:2305.00624, 2023

    2023 arXiv

  34. [42]

    Yaron Lipman, Ricky T. Q. Chen, Haggai Ben-Hamu, Maximilian Nickel, and Matthew Le. flow matching for generative modeling.arxiv preprint arxiv:2210.02747, 2022

    2022 arXiv

  35. [43]

    flow straight and fast: learning to generate and transfer data with rectified flow.arxiv preprint arxiv:2209.03003, 2022

    Yang Liu, Zhen Li, Pranay Kothari, Evangelos Theodorou, and Yang Song. flow straight and fast: learning to generate and transfer data with rectified flow.arxiv preprint arxiv:2209.03003, 2022

    2022 arXiv

  36. [44]

    revisiting classifier two-sample tests

    David Lopez-Paz and Maxime Oquab. revisiting classifier two-sample tests. Ininternational conference on learning representations (iclr), 2017

    2017

  37. [45]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019

    2019

  38. [46]

    Eeg signal reconstruction using a generative adversarial network with a temporal–spatial–frequency loss.Frontiers in Neuroinformatics, 14:15, 2020

    Tian-Jian Luo and Bin Lu. Eeg signal reconstruction using a generative adversarial network with a temporal–spatial–frequency loss.Frontiers in Neuroinformatics, 14:15, 2020

    2020

  39. [47]

    Membership inference attacks against diffusion models.arXiv preprint arXiv:2302.03262, 2023

    Tomoya Matsumoto, Takayuki Miura, and Naoto Yanai. Membership inference attacks against diffusion models.arXiv preprint arXiv:2302.03262, 2023

    2023 arXiv

  40. [48]

    UMAP: Uniform manifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426, 2018

    Leland McInnes, John Healy, and James Melville. UMAP: Uniform manifold approximation and projection for dimension reduction.arXiv preprint arXiv:1802.03426, 2018

    2018 arXiv

  41. [49]

    cGANs with projection discriminator

    Takeru Miyato and Masanori Koyama. cGANs with projection discriminator. InInternational Conference on Learning Representations, 2018

    2018

  42. [50]

    Mognon, J

    A. Mognon, J. Jovicich, L. Bruzzone, and M. Buiatti. Adjust: An automatic eeg artifact detector based on the joint use of spatial and temporal features.Psychophysiology, 48(2):229–240, 2011

    2011

  43. [51]

    Nayak and Sandip Bandyopadhyay

    Chitra S. Nayak and Sandip Bandyopadhyay. mesial temporal lobe epilepsy. InStatPearls. StatPearls Publishing, Treasure Island (FL), 2023. [updated 2023 may 22; cited 2025 jan]

    2023

  44. [52]

    diff-ecg: realistic electrocardiogram generation with diffusion probabilistic models.arxiv preprint arxiv:2306.01875, 2023

    Cherif Neifar, Abdennaceur Kachouri, et al. diff-ecg: realistic electrocardiogram generation with diffusion probabilistic models.arxiv preprint arxiv:2306.01875, 2023

    2023 arXiv

  45. [53]

    improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. improved denoising diffusion probabilistic models. Ininternational conference on machine learning (icml), pages 8162–8171, 2021

    2021

  46. [54]

    Hugh Nolan, Robert Whelan, and Richard B. Reilly. Faster: Fully automated statistical thresholding for eeg artifact rejection.Journal of Neuroscience Methods, 192(1):152–162, 2010. 11

    2010

  47. [55]

    Iden- tifying true brain interaction from eeg data using the imaginary part of coherency.Clinical Neurophysiology, 115(10):2292–2307, 2004

    Guido Nolte, Ou Bai, Lewis Wheaton, Zoltan Mari, Sherry V orbach, and Mark Hallett. Iden- tifying true brain interaction from eeg data using the imaginary part of coherency.Clinical Neurophysiology, 115(10):2292–2307, 2004

    2004

  48. [56]

    The temple university hospital eeg data corpus.Frontiers in Neuroscience, 2016

    Iyad Obeid and Joseph Picone. The temple university hospital eeg data corpus.Frontiers in Neuroscience, 2016

    2016

  49. [57]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas K"opf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu...

    2019

  50. [58]

    Automagic: Standardized preprocessing of big eeg data.bioRxiv, page 460469, 2019

    Andreas Pedroni, Arman Bahreini, and Nicolas Langer. Automagic: Standardized preprocessing of big eeg data.bioRxiv, page 460469, 2019. Preprint

    2019

  51. [59]

    scalable diffusion models with transformers

    William Peebles and Saining Xie. scalable diffusion models with transformers. Ininternational conference on machine learning (icml), 2023

    2023

  52. [60]

    Transformer convolutional neural networks for automated artifact detection in scalp eeg.arXiv preprint arXiv:2208.02405, 2022

    Wei Yan Peh, Yuanyuan Yao, and Justin Dauwels. Transformer convolutional neural networks for automated artifact detection in scalp eeg.arXiv preprint arXiv:2208.02405, 2022

    2022 arXiv

  53. [61]

    film: visual reasoning with a general conditioning layer

    Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. film: visual reasoning with a general conditioning layer. Inaaai conference on artificial intelligence, 2018

    2018

  54. [62]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InMedical Image Computing and Computer-Assisted Inter- vention (MICCAI), 2015

    2015

  55. [63]

    Falk, and Jocelyn Faubert

    Yannick Roy, Hubert Banville, Isabela Albuquerque, Alexandre Gramfort, Tiago H. Falk, and Jocelyn Faubert. Deep learning-based electroencephalography analysis: A systematic review. Journal of Neural Engineering, 16(5):051001, 2019

    2019

  56. [64]

    Mehdi S. M. Sajjadi, Olivier Bachem, Mario Lucic, Olivier Bousquet, and Sylvain Gelly. Assessing generative models via precision and recall. InAdvances in Neural Information Processing Systems, 2018

    2018

  57. [65]

    progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. progressive distillation for fast sampling of diffusion models. arxiv preprint arxiv:2202.00512, 2022

    2022 arXiv

  58. [66]

    Robin Tibor Schirrmeister, Jost Tobias Springenberg, Lukas D. J. Fiederer, Martin Glasstetter, Katharina Eggensperger, Michael Tangermann, Frank Hutter, Wolfram Burgard, and Tonio Ball. Deep learning with convolutional neural networks for eeg decoding and visualization.Human B...

    2017

  59. [67]

    McHugh, Lily Veloso, Meysam Golmoham- madi, Iyad Obeid, and Joseph Picone

    Vinit Shah, Eva von Weltin, Silvia Lopez, James R. McHugh, Lily Veloso, Meysam Golmoham- madi, Iyad Obeid, and Joseph Picone. The temple university hospital seizure detection corpus. Frontiers in Neuroinformatics, 12:83, 2018

    2018

  60. [68]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    2010 arXiv

  61. [69]

    consistency models.arxiv preprint arxiv:2303.01469, 2023

    Yang Song, Chenlin Meng, and Stefano Ermon. consistency models.arxiv preprint arxiv:2303.01469, 2023

    2023 arXiv

  62. [70]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations.arXiv preprint arXiv:2011.13456, 2020

    2011 arXiv

  63. [71]

    evaluation measures for synthetic time series: a taxonomy and empirical study.Journal of Big Data, 11(1):92, 2024

    Philipp Stenger, Paul Krueger, Jan Bauer, et al. evaluation measures for synthetic time series: a taxonomy and empirical study.Journal of Big Data, 11(1):92, 2024. 12

    2024

  64. [72]

    Muhang Tian, Bernie Chen, Allan Guo, Shiyi Jiang, and Anru R. Zhang. Reliable generation of privacy-preserving synthetic electronic health record time series via diffusion models.arXiv, 2024

    2024

  65. [73]

    Dalbagno, and Francesco Fumagalli

    Giulio Tosato, Cesare M. Dalbagno, and Francesco Fumagalli. Eeg synthetic data generation using probabilistic diffusion models.arXiv preprint arXiv:2303.06068, 2023

    2023 arXiv

  66. [74]

    EEG artifact removal—state-of-the-art and guidelines.Journal of Neural Engineering, 12(3):031001, jun 2015

    Jose Antonio Urigüen and Begoña García-Zapirain. EEG artifact removal—state-of-the-art and guidelines.Journal of Neural Engineering, 12(3):031001, jun 2015

    2015

  67. [75]

    Visualizing data using t-SNE.Journal of Machine Learning Research, 9:2579–2605, 2008

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE.Journal of Machine Learning Research, 9:2579–2605, 2008

    2008

  68. [76]

    Julius Vetter, Richard Gao, and Jakob H. Macke. Generating realistic neurophysiological time series with denoising diffusion probabilistic models.Patterns, 5(9):101047, 2024

    2024

  69. [77]

    Peter D. Welch. the use of fast fourier transform for the estimation of power spectra: a method based on time averaging over short, modified periodograms.IEEE Transactions on Audio and Electroacoustics, 15(2):70–73, 1967

    1967

  70. [78]

    tree-ring watermarks: fingerprinting deep generative models.arxiv preprint arxiv:2305.20030, 2023

    Yuxin Wen, Shaofei Li, Micah Goldblum, Avi Schwarzschild, Wojciech Czaja, and Tom Goldstein. tree-ring watermarks: fingerprinting deep generative models.arxiv preprint arxiv:2305.20030, 2023

    2023 arXiv

  71. [79]

    Automatic classification of artifactual ica-components for artifact removal in eeg signals.Behavioral and Brain Functions, 7(30):1–15, 2011

    Irene Winkler, Stefan Haufe, and Michael Tangermann. Automatic classification of artifactual ica-components for artifact removal in eeg signals.Behavioral and Brain Functions, 7(30):1–15, 2011

    2011

  72. [80]

    parallel wavegan: a fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram

    Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. parallel wavegan: a fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. ieee/acm transactions on audio, speech, and language processing, 28:1837–1847, 2020

    2020

  73. [81]

    Time-series generative adversarial networks

    Jinsung Yoon, Daniel Jarrett, and Mihaela van der Schaar. Time-series generative adversarial networks. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  74. [82]

    adding conditional control to text-to-image diffusion models.arxiv preprint arxiv:2302.05543, 2023

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. adding conditional control to text-to-image diffusion models.arxiv preprint arxiv:2302.05543, 2023. 13 A Appendices and Supplementary Material A.1 Compute & Environment All experiments were run on a single workstation; we provide ex...

    2023 arXiv

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.