Pith. sign in

REVIEW 1 major objections 6 minor 16 references

Generative Augmentation for EEG Motor Imagery Classification: A Class-Conditional VAE with Cycle-Consistent Decoder Refinement

T0 review · 1 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read The paper claims that synthetic EEG from a class-conditional VAE is best understood as class-structured, covariance-like data rather than a substitute for real EEG: it can lift the MDM point estimate, but augmentation gains are small and cl

desk verdict An honestly scoped, well-reported negative result: this CVAE learns class structure but doesn't reliably boost augmentation accuracy, and the null is underpowered enough (3 seeds, 4 subjects, one dataset) that the paper's value is in its calibration and ablations, not a definitive verdict. read the letter →

arxiv 2607.22733 v1 pith:2YHHRKII submitted 2026-07-22 cs.CV cs.LG

Matei Moldoveanu , Alain Sirois , Claire Ben Ali , Fabien Lotte , Florian Yger This is my paper · ORCID
classification cs.CVcs.LG
keywords EEGmotorimagerybrain-computerinterfacedataaugmentationclass-conditionalvariationalautoencoderSoft-DTWLog-EuclideancovarianceRiemanniangeometry
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 a generative model can manufacture useful synthetic motor-imagery EEG trials that improve downstream classifiers. It trains a class-conditional variational autoencoder with a learned per-class latent prior, a latent classifier for class separation, a Log-Euclidean covariance constraint on decoded signals, and a decoder-refinement phase that cycles prior samples through the encoder. Across four standard classifiers and two evaluation regimes, adding synthetic trials produced no statistically reliable accuracy gains; only a minimum-distance Riemannian classifier showed a positive point-estimate trend. The authors' conclusion is that the generator's value lies in producing class-structured, covariance-consistent synthetic data for probing what classifiers extract from class-level structure, not in replacing or supplementing real EEG. A fair reading is that the augmentation claim is deliberately conservative, with the main open question being statistical power.

What carries the argument

The carrying mechanism is a class-conditional variational autoencoder whose decoder is reused as the generator: after training, one samples the learned per-class Gaussian prior in a 100-dimensional latent space and decodes it into a synthetic 5-channel, 500-sample trial. Three elements make the model work as intended: a Soft-DTW reconstruction loss that is tolerant of small temporal shifts, a Log-Euclidean penalty that matches the 5x5 channel-covariance matrix of the decoded signal to the real trial, and a latent classifier plus conditional KL term that organizes the latent space by class. A fourth element, the cycle-consistent decoder refinement, alternates ordinary VAE training with decode

What would settle it

Run the same augmentation protocol with substantially more seeds (for example, 30), more held-out subjects, and at least one additional motor-imagery dataset with a different channel montage. If the mean with-augmentation delta across the four classifiers produces a 95% confidence interval that excludes zero in either protocol, the paper's central conservative claim is falsified. Alternatively, a positive control—adding real re-sampled trials or a known-good augmentation method and showing that the protocol detects it—would test whether the null result is a property of the generator or of the

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a class-conditional VAE trained on motor-imagery EEG learns a latent space with real class structure—held-out trials can be classified by nearest learned prior mean at roughly 0.8 accuracy, and classifiers trained only on generated trials reach well above chance for covariance-based pipelines—yet this structure does not translate into reliable gains when synthetic trials are added to real training data. Across within-user and cross-user protocols, the with-augmentation deltas for CSP+LDA, tangent-space SVM, and EEGNet are small and their 95% confidence intervals include zero; the minimum-distance Riemannian classifier is the only pipeli

Load-bearing premise

The load-bearing premise is that the experimental design—three random seeds, four leave-one-subject-out folds, and one five-channel dataset—is powerful enough to detect the augmentation effect it is trying to measure; if the design is underpowered, as the authors themselves note, the absence of significant gains cannot be read as evidence that no benefit exists.

Editorial extensions

If this is right

  • For the four classifiers tested, adding a training-sized batch of synthetic trials does not produce accuracy gains with confidence intervals excluding zero, under either within-user or cross-user protocols.
  • The augmentation volume sweep from 0.5x to 5x does not reveal a reliable benefit for any classifier; MDM's point estimate trends upward within-user and plateaus around 1x, but the interval is wide.
  • Synthetic-only training reaches well above chance for covariance-based pipelines but degrades EEGNet sharply, indicating the generated data carries class-level covariance structure but not the temporal detail neural networks need.
  • The learned latent space is discriminative enough that nearest-prior-mean classification matches signal-space pipelines, but that discriminability adds little information beyond the real training set, so better decoders alone would not fix augmentation.

Reading between the lines

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

  • Editorial extension: the same generated samples could be used deliberately as a diagnostic set—class-prototypical, covariance-preserving probes—to measure how much of a pipeline's accuracy comes from covariance structure rather than temporal detail; the paper does not propose this use.
  • Editorial extension: a direct test would be to use the generated covariance matrices to regularize Riemannian covariance estimates for low-data classes, or to interpolate between real and synthetic SPD matrices; the paper's setup does not evaluate this.
  • Editorial extension: because removing the covariance constraint also collapses the encoder's latent class discrimination, the results suggest the model's usefulness is bounded by the extent to which discriminative EEG variability is covariance-related; a task where class differences live more in waveform detail would likely show even less augmentation benefit.
  • Editorial extension: a larger study—more seeds, more subjects, more datasets—could overturn the conservative conclusion, and the paper's own limitation section explicitly invites that check.
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

1 major / 6 minor

Summary. The paper proposes a class-conditional VAE with an integrated latent classifier, a Log-Euclidean covariance constraint, and an alternating decoder-focused cycle-consistency refinement for generating synthetic motor-imagery EEG. It evaluates augmentation by adding synthetic trials to the training sets of four downstream classifiers (CSP+LDA, TGSP+SVM, MDM, EEGNet) under within-user (3 seeds) and cross-user (LOSO, 4 subjects) protocols on the Zhou2016 dataset. The central finding is that the generator captures class-structured, covariance-like information (nearest-prior-mean accuracy ~0.8; MDM Aug-only >0.7) but that adding synthetic trials does not significantly improve accuracy in most conditions, with MDM showing only modest, non-significant positive point estimates. The paper includes extensive ablations and per-seed/per-subject tables in the appendix.

Significance. If the empirical claims are taken at face value, the paper provides a useful qualified result for EEG augmentation: a covariance-regularized CVAE can produce class-structured synthetic data that is useful for probing classifiers but not a reliable augmentation source. The paper is methodologically careful in several respects: it reports confidence intervals, discloses limitations, performs seven ablations, and provides per-seed/per-subject tables allowing readers to assess variability. The central negative result is genuinely informative for the BCI community, where augmentation claims are often optimistic. However, the significance is bounded by the small scale (single dataset, 3 seeds, 4 folds) and the fact that the main conclusion is an absence claim supported by null results from a small sample.

major comments (1)
  1. [Abstract and §5.1] The abstract states that synthetic EEG 'can raise the point estimate for MDM.' In Table 6, MDM with-Aug delta is +0.014 ± 0.024 and Aug-only +0.045 ± 0.093; in Table 9 cross-user, with-Aug delta is +0.009 ± 0.053 and Aug-only +0.013 ± 0.121. None of these exclude zero. The paper is appropriately careful in the body, but the abstract's phrasing is potentially misleading—a point estimate that is not statistically distinguishable from zero is not a 'credible' benefit without a power analysis. I recommend rewording to 'MDM showed positive point estimates in both protocols, but these were not statistically significant in this sample.'
minor comments (6)
  1. [§3.3, Eq. (5)] The loss weights are reported as constants; it would be helpful to state whether they were tuned on a validation set or chosen a priori, and whether the ablations in §A.5–A.7 informed the final choice or were after-the-fact.
  2. [§4.2, Table 4] For CSP+LDA, the number of CSP components (8) and LDA regularization are not specified; for TGSP+SVM, the SVM C and RBF gamma are not given. Given the paper's reproducibility emphasis, these hyperparameters should be listed, even if they are library defaults.
  3. [§5.3.1, Figure 3] The caption says 'latent posteriors for train, validation and test trials, the directly-sampled class-conditional prior ("aug"), and the per-class prior means.' It is unclear whether the 'aug' points are drawn from the prior or are posterior encodings of generated samples; please clarify in the caption or text.
  4. [§5.3.2, Table 13] Soft-DTW distances are reported without normalization details (e.g., whether normalized by trial length or channel count). Raw distances of ~1883 vs 283 are hard to interpret; consider normalizing or providing units.
  5. [Appendix A.1, Table 15] The ablation table reports only point estimates (no CI) and uses a single seed, but the text interprets the alternating schedule as 'bringing a large improvement to MDM.' Given the paper's own emphasis on seed variability, this claim should be hedged or accompanied by additional seeds.
  6. [§4.1, within-user split] The within-user protocol pools across subjects and sessions, meaning the test set contains trials from the same subjects used in training. This is standard for within-subject evaluation but should be explicitly noted as a potential source of optimistic accuracy.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity: the augmentation evaluation is held-out and independent; only a minor by-construction covariance diagnostic.

  1. self definitional [Eq. (3) (Lcov) and Sec. 5.3.4 / Fig. 8]
    "Lcov = || log Σ(ˆx) − log Σ(x) ||_F^2 (3) ... 'Generated maps otherwise show broadly similar behaviour to the train data on the two better-behaved classes, consistent with the covariance constraint improving part of the spatial fidelity of the generated data.'"

    The abstract/conclusion claim that the CVAE is 'most credible as a source of class-structured, covariance-like data' is partly supported by reporting that generated per-channel variance correlates with real train data (0.909 for feet). But Eq. (3) directly minimizes the log-Euclidean distance between the covariance of the decoded output and the real trial covariance. Observing covariance similarity to the training set is therefore a restatement of the fitted objective, not an independent confirmation. This is minor and non-load-bearing: the central negative augmentation result rests on held-out test accuracy, not on this diagnostic.

full rationale

The paper's claimed derivation chain is: train a CVAE on real MI-EEG trials using reconstruction, KL, covariance-matching and classification losses; sample the learned class-conditional prior; decode into synthetic trials; train four independent downstream classifiers on real, real+synthetic, or synthetic-only data; score on held-out test partitions (within-user 60/20/20, cross-user LOSO with the target subject withheld). None of the central augmentation deltas is fitted to the test set: z-scoring uses training statistics only, the CVAE never sees the test subject in LOSO, and the baseline/with-Aug differences are measured accuracies, not identities. The negative main result (deltas with CIs including zero) is therefore an empirical null, not an analytic consequence of the training objective. There is no load-bearing self-citation: references to RGP-VAE, Doku et al., CVAE, Soft-DTW, and RAVE are external prior work with no author overlap with the present paper, and no 'uniqueness theorem' or ansatz is imported from the authors' own prior publications. The latent-space and covariance diagnostics are partially aligned with the model's own losses (Lcls, LKL, Lcov), which is why one minor by-construction step is noted for the 'covariance-like' characterization, but the paper explicitly attributes this to the covariance constraint and does not use it to derive augmentation benefit. Section 7 flags the study as 'likely underpowered for some comparisons'; that is a statistical-power concern for the absence claim, not a circularity. Overall, no significant circularity; score 1 reflects only the minor, non-load-bearing diagnostic tautology.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical entities or forces. The trainable class-conditional prior and latent classifier are model components, not independent postulates. The central claims rest on a small number of hand-chosen hyperparameters (loss weights, latent dim, augmentation volume) and on domain assumptions about the EEG dataset and covariance representation. No formal verification or fully independent validation is provided.

free parameters (5)
  • Loss weights (lambda_rec=0.01, lambda_kl=5.0, lambda_cov=1.5, lambda_cls=1.0) = 0.01, 5.0, 1.5, 1.0
    Hand-chosen and tuned via ablation studies on the same dataset; affects the tradeoff between reconstruction, latent separation, and covariance fidelity. Eq. 5; Appendices A.5-A.7.
  • Latent dimension = 100
    Architecture choice; no principled justification beyond empirical performance. Section 3.2.
  • Soft-DTW gamma = 1
    Set to 1 without a sensitivity sweep; this controls the smoothness of the DTW alignment in Eq. 1.
  • Covariance regularization epsilon = 1e-6
    Added to the sample covariance matrix to make it strictly SPD; Section 3.3.
  • Augmentation volume 1x = equal to training set size
    Main results use a synthetic set the same size as the real training set; volumes 0.5x-5x are explored, but the choice of anchor volume is arbitrary. Section 4.3.
assumptions (5)
  • domain assumption Zhou2016 EEG trials are correctly labelled as feet/left-hand/right-hand and are representative of motor-imagery EEG.
    All training, generation, and evaluation depend on label correctness and dataset representativeness; stated in Section 4.1.
  • domain assumption The 5x5 sample channel covariance matrix regularized by epsilon*I is a sufficient statistic for the spatial structure used by downstream classifiers.
    The Log-Euclidean covariance loss and the Riemannian classifiers (MDM, TGSP+SVM) rely on this; Section 3.3 Eq. 3.
  • domain assumption Soft-DTW distance with gamma=1 is an appropriate reconstruction objective for EEG trials.
    The model uses it as the reconstruction loss; Section 3.3 Eq. 1 and ablation A.2.
  • standard math Log-Euclidean distance is a valid metric on the manifold of SPD matrices.
    Used in Eq. 3 to compare covariance matrices; standard result in Riemannian geometry.
  • ad hoc to paper A 100-dimensional Gaussian latent space with trainable per-class priors can capture class-discriminative EEG structure.
    This is the model's core representational assumption; no independent justification beyond the empirical results. Section 3.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative Augmentation for EEG Motor Imagery Classification: A Class-Conditional VAE with Cycle-Consistent Decoder Refinement." pith.science (2026). https://pith.science/paper/2YHHRKII

@misc{pith2026260722733,
  author       = {Pith},
  title        = {Pith review of: Generative Augmentation for EEG Motor Imagery Classification: A Class-Conditional VAE with Cycle-Consistent Decoder Refinement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2YHHRKII}},
  note         = {Machine review of arXiv:2607.22733}
}
abstract

We investigate whether a generative model can supply useful synthetic motor-imagery (MI) electroencephalography (EEG) trials that improve the accuracy of independent downstream classifiers. We train a class-conditional variational autoencoder (CVAE) with an integrated latent classifier on the Zhou motor-imagery dataset, using the learned per-class prior as a generator: sampling the prior for a given label and decoding it into a synthetic, label-consistent signal. A constraint on the covariance matrix of the generated data encourages preservation of covariance structure, and the model is trained with a schedule that alternates ordinary VAE training with a decoder-focused phase that sharpens the generative pathway used for augmentation. We measure the effect of adding synthetic trials to the training set under two evaluation protocols -- within-user (pooled 60/20/20 split across subjects) and cross-user (leave-one-subject-out, LOSO) -- across four representative EEG classification pipelines: Common Spatial Patterns with Linear Discriminant Analysis (CSP+LDA), tangent-space features with a Support Vector Machine (TGSP+SVM), Minimum Distance to Riemannian Mean (MDM), and a neural network based on EEGNetv4 (henceforth EEGNet). Results are aggregated across independent augmentation draws, random seeds (within-user), or leave-one-subject-out folds (cross-user), with uncertainty reported as 95\% confidence intervals (Student's $t$-distribution) computed over per-seed/per-fold averages. We find that synthetic EEG from the CVAE is most credible as a source of class-structured, covariance-like data rather than as a substitute for real raw EEG: it can raise the point estimate for MDM, but the broader augmentation claim remains conservative -- observed gains are small and classifier-dependent.

Figures

Figures reproduced from arXiv: 2607.22733 by the authors.

Figure 1
Figure 1. Model architecture. The encoder maps a trial to a Gaussian posterior; a latent sample is decoded and classified. The class-conditional prior is trainable per class. Trials are 3-class (feet / left-hand / right-hand), 5 channels, 500 time samples (5 s at 100 Hz), 4 [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Decoder-focused phase, prior→decode→re-encode cycle. Encoder, prior, and classifier are frozen (grey); only the decoder (green) receives gradient updates. In the matched 3-seed ablation reported in Appendix A.1, the alternating schedule was asso￾ciated with improved point estimates for the usefulness of generated data for augmentation: it moved MDM’s with-Aug delta from net-negative to net-positive and raised the ac… view at source ↗
Figure 3
Figure 3. Latent space (t-SNE). Latent posteriors for train, validation and test trials, the directly-sampled class-conditional prior (“aug”), and the per-class prior means (stars) [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (10 more)
Figure 4
Figure 4. Figure 4: Posterior reconstructions (top: train trials; bottom: test trials; real in pink, reconstruction in orange; rows are channels) [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: t-SNE embedding of pairwise Soft-DTW distance, 20 real feet trials vs. their 20 posterior recon￾structions (train split, seed 0; circles = real, triangles = reconstructions; grey lines connect each real trial to its own reconstruction). Distances are the normalised Sof…
Figure 6
Figure 6. Figure 6: Single-sample FFT (one real feet trial vs. one generated feet sample; per channel; not averaged). 5.3.4 Topographic Maps Topographic maps of per-channel variance, evaluating the spatial covariance structure of each class, are also shown. With only five channels, the ma…
Figure 7
Figure 7. Figure 7: Trial-averaged spectra with variability shading, real vs. generated — feet (one panel per channel; solid line = mean across trials; shaded band = ±1 SD; real in pink, generated in olive) [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Per-channel variance topomaps, feet (columns: train/test/generated). Train-vs-test correlation 0.971 (real data’s own split-to-split consistency), train-vs-generated correlation 0.909, generated total vari￾ance = 77% of train [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Per-channel variance topomaps, left_hand (columns: train/test/generated). Train-vs-test cor￾relation is only 0.379 here — well below feet’s 0.971 — meaning even the real data’s own train/test split is spatially inconsistent for this class, before the generator is invol…
Figure 10
Figure 10. Figure 10: Per-channel variance topomaps, right_hand (columns: train/test/generated). Train-vs-test correlation 0.976 and train-vs-generated correlation 0.865 — both close to feet’s numbers, and generated total variance (95% of train) is the closest to real of the three classes.…
Figure 11
Figure 11. Figure 11: Trial-averaged FFT amplitude, real (pink) vs. reconstruction (olive), MSE (top) vs. Soft-DTW (bottom), all 5 channels. This makes the difference precise: MSE’s reconstruction spectrum peaks near 0–5 Hz, with little to no power in the real mu-band peak (≈10–15 Hz). Sof…
Figure 12
Figure 12. Figure 12: Latent space (t-SNE), KL-only variant. Same construction as [PITH_FULL_IMAGE:figures/full_fig_p023_12.png]
Figure 13
Figure 13. Figure 13: Latent space (t-SNE), both KL and classifier (reported variant). A.4 Covariance Constraint — On vs. Off Looking at the impact for Log-Euclidean covariance, everything else matched. Normal schedule, 1 seed, 200 epochs [PITH_FULL_IMAGE:figures/full_fig_p024_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 4 canonical work pages

  1. [1]

    Learning structured output representation using deep conditional generative models

    Kihyuk Sohn, Xinchen Yan, and Honglak Lee. Learning structured output representation using deep conditional generative models. In Advances in Neural Information Processing Systems (NeurIPS), volume 28, pages 3483–3491, 2015

    2015

  2. [2]

    de Jong, and Andreea I

    Valters Poļaka, Ivo P. de Jong, and Andreea I. Sburlea. Riemannian geometry-preserving variational autoencoder for MI-BCI data augmentation, 2026

    2026

  3. [3]

    Chuang, Ozan Özdenizci, and Albrecht Schmidt

    David Bethge, Philipp Hallgarten, Tobias Grosse-Puppendahl, Mohamed Kari, Lewis L. Chuang, Ozan Özdenizci, and Albrecht Schmidt. EEG2Vec: Learning affective EEG repre- sentations via variational autoencoders. In Proc. 2022 IEEE Int. Conf. Systems, Man, and Cybernetics (SMC) , pages 3150–3157, 2022. doi: 10.1109/SMC53654.2022.9945517

    arXiv 2022

  4. [4]

    Alberto Zancanaro, Giulia Cisotto, Italo Zoppis, and Sara L. Manzoni. vEEGNet: Learning la- tent representations to reconstruct EEG raw data via variational autoencoders. In Information and Communication Technologies for Ageing Well and e-Health , Communications in Computer and Information Science, pages 114–129. Springer, 2024. doi: 10.1007/978-3-031-62753-8_7

    work page doi:10.1007/978-3-031-62753-8_7 2024

  5. [5]

    Zoppis, and Sara L

    Giulia Cisotto, Alberto Zancanaro, Italo F. Zoppis, and Sara L. Manzoni. hvEEGNet: A novel deep learning model for high-fidelity EEG reconstruction. Frontiers in Neuroinformatics , 18: 1459970, 2024. doi: 10.3389/fninf.2024.1459970

    arXiv 2024

  6. [6]

    V AEEG: Variational auto-encoder for extracting EEG representation

    Taorong Zhao, Yuxin Cui, Ting Ji, Jinyuan Luo, Wenlong Li, Jian Jiang, Zheng Gao, Wei Hu, Yueming Yan, Yuwu Jiang, and Bo Hong. V AEEG: Variational auto-encoder for extracting EEG representation. NeuroImage, 304:120946, 2024. doi: 10.1016/j.neuroimage.2024.120946

    arXiv 2024

  7. [7]

    Structure-preserving EEG augmentation via Riemannian conditional generative adversarial networks

    Michael Doku, Imad Eddine Tibermacine, Stefania Russo, Abdelkader Rabehi, Maad Habib, and Christian Napoli. Structure-preserving EEG augmentation via Riemannian conditional generative adversarial networks. IEEE Access, 2026. doi: 10.1109/ACCESS.2026.3663245

    arXiv 2026

  8. [8]

    Williams, Damian Weinhardt, Jasper Hewson, Martyna B

    Christoph C. Williams, Damian Weinhardt, Jasper Hewson, Martyna B. Płomecka, Nicolas Langer, and Sebastian Musslick. EEG-GAN: A generative EEG augmentation toolkit for enhancing neural classification. bioRxiv, 2025. doi: 10.1101/2025.06.23.661164

    work page doi:10.1101/2025.06.23.661164 2025

Show all 16 references
  1. [9]

    Trans-cV AE-GAN: Transformer-based cV AE-GAN for high-fidelity EEG signal generation

    Yuang Yao, Xin Wang, Xiaoyu Hao, Han Sun, Rui Dong, and Yang Li. Trans-cV AE-GAN: Transformer-based cV AE-GAN for high-fidelity EEG signal generation. Bioengineering, 12 (10):1028, 2025. doi: 10.3390/bioengineering12101028

    2025 doi

  2. [10]

    Soft-DTW: a differentiable loss function for time-series

    Marco Cuturi and Mathieu Blondel. Soft-DTW: a differentiable loss function for time-series. In Proc. 34th Int. Conf. Machine Learning (ICML) , volume 70 of PMLR, pages 894–903, 2017. arXiv:1703.01541. 19

    2017 arXiv

  3. [11]

    Ananya Harsh Jha, Saket Anand, Maneesh Singh, and V. S. R. Veeravasarapu. Disentangling factors of variation with cycle-consistent variational auto-encoders. In Computer Vision – ECCV 2018 , volume 11207 of Lecture Notes in Computer Science , pages 829–845. Springer,

    2018

  4. [12]

    CoV AE: Consistency training of variational autoen- coders, 2025

    Gabriele Silvestri and Luca Ambrogioni. CoV AE: Consistency training of variational autoen- coders, 2025

    2025

  5. [13]

    RA VE: A variational autoencoder for fast and high- quality neural audio synthesis, 2021

    Antoine Caillon and Philippe Esling. RA VE: A variational autoencoder for fast and high- quality neural audio synthesis, 2021

    2021

  6. [14]

    A fully automated trial selection method for optimization of motor imagery based brain-computer interface

    Bin Zhou, Xiaopei Wu, Zhen Lv, Lai Zhang, and Xin Guo. A fully automated trial selection method for optimization of motor imagery based brain-computer interface. PLOS ONE , 11(9): e0162657, 2016. doi: 10.1371/journal.pone.0162657

    2016 doi

  7. [15]

    Visualizing data using t-SNE

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of Machine Learning Research, 9(86):2579–2605, 2008. 20 A Appendix A: Ablation Studies These single-factor studies isolate the contribution of individual design choices; each varies only the fact...

    2008

  8. [2018]

    doi: 10.1007/978-3-030-01219-9_49

    doi

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.