Pith. sign in

REVIEW 4 major objections 4 minor 52 references

One-shot handwriting GAN cuts FID below half of the next-best model

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

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

FW-GAN combines wave-modulated MLP generation with frequency-domain losses and a wavelet-based discriminator to synthesize handwriting from one example, reporting lower FID scores than prior methods on IAM and HANDS-VNOnDB.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection Real architectural contribution undercut by an overclaimed headline: the big FID gap is same-word reconstruction, and the true one-shot gain is small and unquantified. the 4 major comments →

arxiv 2508.21040 v1 pith:MKMCWYE3 submitted 2025-08-28 cs.CV cs.LG

FW-GAN: Frequency-Driven Handwriting Synthesis with Wave-Modulated MLP Generator

Huynh Tong Dang Khoa , Dang Hoai Nam , Vo Nguyen Le Duy This is my paper
classification cs.CV cs.LG
keywords handwriting synthesisgenerative adversarial networksfrequency domainwavelet transformone-shot learningWave-MLPhandwritten text recognitiondata augmentation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

The paper claims that handwriting synthesis improves when the generator is explicitly aware of frequency content, not just pixels. It proposes FW-GAN, a one-shot GAN that renders arbitrary words in a writer's style from a single reference image. The generator inserts phase-aware Wave-MLP blocks into a BigGAN-style backbone, a second discriminator judges high-frequency wavelet coefficients, and a frequency distribution loss aligns spectral statistics between real and generated handwriting. On English (IAM) and Vietnamese (HANDS-VNOnDB) benchmarks, the model reports lower FID/KID than HWT, VATr, HiGAN, and HiGAN+, and synthetic data from it lowers recognition error in a low-resource HTR setup. If the comparison protocol is fair, frequency-aware architectures would be a practical default for handwriting data augmentation.

Core claim

FW-GAN is a one-shot adversarial framework for offline handwriting imitation. Given one word image by a writer and an arbitrary character sequence, a hierarchical generator renders the sequence in that writer's style, with the style either sampled from a prior or encoded from the reference. The paper's central claim is that modeling phase and amplitude, plus explicit supervision on high-frequency content, is what makes the synthesis realistic. Three additions carry the argument: Phase-Aware Token Mixing treats each spatial token as a complex wave with a learned phase that modulates token mixing; a dual discriminator pairs a standard spatial discriminator with one that runs on Haar-wavelet de

What carries the argument

The key component is the WaveGBlock, a generator block that fuses a conditional style vector with character content and uses Phase-Aware Token Mixing (PATM), where each spatial token is represented by amplitude and a learned phase and mixed along height and width as complex waves. It is supported by a dual discriminator: the standard spatial discriminator plus a high-frequency discriminator that applies a fixed 2D Haar wavelet transform and scores the summed LH, HL, and HH detail bands. A third load-bearing piece is the Frequency Distribution Loss, which computes the sliced Wasserstein distance between amplitude and phase spectra of deep features extracted from real and generated images. Tog

Load-bearing premise

The evaluation assumes that the from-scratch re-implementations of HiGAN and HiGAN+ are as strong as the original published models; if those baselines were undertuned, the reported FID advantage would be overstated.

What would settle it

Run the original released HiGAN and HiGAN+ checkpoints on the same IAM train/test split and compute FID and KID on the same generated sample sets. If those released models match or beat FW-GAN's reported FID of 6.53, the central claim fails; at minimum, showing that the paper's re-implementations reproduce the originally published FID numbers would settle the fairness of the comparison.

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

If this is right

  • If the claims hold, frequency-aware discriminators and spectral alignment losses become a plug-in improvement for GAN-based handwriting generation, not just a benchmark trick.
  • Synthetic data from FW-GAN can lower character and word error rates in low-resource OCR, so the model can substitute for some real labeled handwriting in training.
  • One-shot style conditioning means a single reference word suffices to render out-of-vocabulary words in a novel writer's style, which is useful for personalization and archival reconstruction.
  • Wave-MLP blocks offer a lighter alternative to transformer backbones for long-range dependencies, with the reported model size at 52.96 MB versus 131–155 MB for the transformer baselines.

Where Pith is reading between the lines

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

  • The frequency loss and wavelet discriminator are not handwriting-specific; a testable extension is applying the same dual-discriminator plus spectral distribution loss to other structured image generation tasks such as signatures, music notation, or diagrams where stroke sharpness matters.
  • The reported gains may partly depend on the from-scratch re-implementations of HiGAN and HiGAN+; a direct checkpoint-to-checkpoint replication on the same split would clarify how much of the improvement is architectural.
  • Because the frequency distribution loss operates on deep features rather than pixels, it may offer a generic fix for spatial misalignment in style transfer beyond handwriting, for example in font or glyph generation.
  • The KL-regularized style space allows sampling novel styles, suggesting the model could be used for writer-style interpolation or for synthesizing entirely new writers, an application the paper does not explore.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes FW-GAN, a GAN framework for one-shot handwriting synthesis. The generator combines a BigGAN-style backbone with Wave-MLP/PATM modules; training adds a Frequency Distribution Loss (FDL), a high-frequency discriminator based on Haar wavelet decomposition, and auxiliary recognition/writer/style/KL losses. Experiments on IAM (English) and HANDS-VNOnDB (Vietnamese) compare against HiGAN, HiGAN+, HWT, and VATr in five evaluation scenarios, reporting FID/KID, downstream HTR augmentation results, an ablation, model sizes, and qualitative examples. The paper claims FW-GAN 'significantly outperforms the state-of-the-art'.

Significance. If the empirical claims hold, FW-GAN is a useful contribution to offline handwriting synthesis and to low-resource HTR augmentation. The paper has clear strengths: it defines multiple evaluation scenarios (IV/OOV and seen/unseen style), reports downstream HTR metrics, includes a component ablation, tests cross-language generalization, and provides a public implementation URL. However, the evidence for the strongest claims is weakened by the evaluation protocol and the lack of any uncertainty quantification, and the claim that FDL is 'novel' is internally inconsistent with the citation to prior work.

major comments (4)
  1. [Section 4.1, Tables 1 and 2] The abstract's 'significantly outperforms the state-of-the-art' is supported mainly by the Test Set Replication condition in Table 1, where the style reference is the same word as the target ('the style is derived from reference images of the same writer containing the same word'). This is close to reconstruction rather than one-shot synthesis of arbitrary text. In the genuinely one-shot condition, OOV-U in Table 2, FW-GAN's advantage over the official-checkpoint baselines HWT and VATr is only 0.88 and 0.70 FID points (28.80 vs 29.68 and 29.50). None of the tables report standard deviations, repeated runs, or significance tests. The phrase 'significantly outperforms' is therefore not supported by the hardest and most relevant condition, even if every baseline were perfectly tuned. Please either restrict the claim to the replication and in-vocabulary settings, or provide uncertainty estim
  2. [Section 4.3] The comparison protocol is asymmetric: HWT and VATr use released checkpoints, while HiGAN and HiGAN+ are re-implemented from scratch by the authors. The paper states this is done to ensure identical data partitions, but it gives no evidence that the re-implementations reproduce the originally published HiGAN/HiGAN+ results or match their hyperparameters. If the re-implementations are suboptimal, the reported gains over HiGAN/HiGAN+ (e.g., 17.086 vs 6.530 in Table 1) are inflated. Please validate the re-implementations against published numbers on the same split, or use the official checkpoints/code for all baselines.
  3. [Section 1 and Section 3.3.6] The paper claims as a contribution 'a novel loss function' and an abstract 'novel Frequency Distribution Loss', but Section 3.3.6 states 'we adopt the Frequency Distribution Loss (FDL) [9]', citing Ni et al. (CVPR 2024). This is internally inconsistent and overstates the novelty. Please reframe the contribution as integrating/adapting FDL into handwriting synthesis, or identify a concrete modification beyond [9] (e.g., specific feature layer, amplitude/phase weighting, or combination with the high-frequency discriminator).
  4. [Tables 3, 4, 5; Sections 4.4-4.6] All reported metrics are single-run point estimates, with no error bars or significance tests. This is especially important where the reported margins are small, e.g., Table 3 WER 28.18 vs 29.65 for VATr, and Table 5 FID 5.61 vs 9.85 for HWT. FID/KID are Monte Carlo estimates and CER/WER depend on a single HTR training run. To support 'consistent superiority' and the practical utility claim, please provide at least three independent training/evaluation runs with mean and standard deviation, or paired statistical tests (e.g., bootstrap over generated samples) for all headline comparisons.
minor comments (4)
  1. [Section 4.2] The adaptive balancing strategy for lambda_R, lambda_W, and lambda_style is described only as 'adaptively tuned throughout training using a gradient balancing strategy'. Please specify the algorithm, update rule, and final values so the training objective in Eq. (20) is reproducible.
  2. [Section 3.3.6, Eqs. (16)-(17)] The notation is overloaded: A denotes both the query set in Section 3.1 and a feature set in Eq. (16), and A(A) / P(A) are hard to parse. Please rename the feature sets (e.g., F_x, F_g) and define the amplitude/phase extraction from the DFT explicitly.
  3. [Section 4.6, Table 5] The Vietnamese evaluation is said to follow 'the same experimental setup described in 4.2', but Table 5 reports only one FID/KID value. Please state which of the five scenarios is used (e.g., Test Set Replication or OOV-U) and the number of generated samples.
  4. [References] Reference [7] contains a corrupted author string ('Z.-Q. J. X. Zhi-Qin John Xu, Y. Z. Yaoyu Zhang...'). Please clean the reference list and ensure all entries use standard author/venue formatting.

Circularity Check

0 steps flagged

No significant circularity: empirical benchmarking against external datasets and baselines, with no self-citation or fitted-input-called-prediction chain.

full rationale

FW-GAN is an empirical systems paper. Its central claims are evaluated with FID/KID against the IAM and HANDS-VNOnDB benchmarks and with downstream HTR metrics (CER/NED/WER). All comparisons are to external baselines (HiGAN, HiGAN+, HWT, VATr), and the proposed components are either introduced in this paper (high-frequency discriminator) or adopted from external prior work with citations (Wave-MLP [8], Frequency Distribution Loss [9]). No equation defines the predicted quantity in terms of itself, and no fitted parameter is renamed as a prediction: the model is trained on the training split and evaluated on held-out writers/words. The only limitations noted in the manuscript—e.g., the writer classifier may not generalize to unseen writers (Sec. 3.3.3), and the same-word Test Set Replication protocol makes that condition reconstruction-like—are evaluation-protocol and generalization concerns, not circularity. There is no self-citation, no imported uniqueness theorem, and no load-bearing appeal to the authors' own prior work. The possibility that the HiGAN/HiGAN+ re-implementations underperform the original tuned models is a fairness concern, not a circularity concern, because the comparison is still external and empirical.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions for GAN training and on the validity of FID/KID for small handwriting images. No new physical entities are introduced; the wave-token representation is from Wave-MLP [8]. The loss-balancing hyperparameters are hand-set or adaptively tuned, so they are recorded as free parameters.

free parameters (1)
  • Loss weighting coefficients (lambda_R, lambda_W, lambda_style, lambda_FDL, lambda_KL, lambda in FDL) = lambda_FDL=1, lambda_KL=0.0001; others adaptively tuned
    Hand-chosen or adaptively tuned to balance competing objectives; the paper fixes two and leaves the rest to a gradient balancing strategy.
axioms (4)
  • domain assumption The F-Principle (neural networks learn low frequencies first) motivates frequency-based supervision.
    Sections 1 and 2; this motivates the frequency losses but is not proven for GAN training in this setting.
  • domain assumption Writer identity is a usable proxy for handwriting style and the style space is approximately Gaussian.
    Sections 3.3.3 and 3.3.5; the writer classifier and KL loss assume style is class-discriminable and unimodal Gaussian.
  • domain assumption FID/KID computed with Inception features are valid quality metrics for 32x128 handwriting images.
    Section 4.1; this is an accepted practice in the cited literature, but FID on tiny document images is sensitive to preprocessing.
  • domain assumption The recognizer R trained on real data only transfers as a content supervisor to generated images.
    Section 3.3.2; the frozen recognizer's CTC gradient is assumed to enforce legibility without being fooled by distribution shift.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of FW-GAN: Frequency-Driven Handwriting Synthesis with Wave-Modulated MLP Generator." pith.science (2026). https://pith.science/paper/MKMCWYE3

@misc{pith2026250821040,
  author       = {Pith},
  title        = {Pith review of: FW-GAN: Frequency-Driven Handwriting Synthesis with Wave-Modulated MLP Generator},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MKMCWYE3}},
  note         = {Machine review of arXiv:2508.21040}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Labeled handwriting data is often scarce, limiting the effectiveness of recognition systems that require diverse, style-consistent training samples. Handwriting synthesis offers a promising solution by generating artificial data to augment training. However, current methods face two major limitations. First, most are built on conventional convolutional architectures, which struggle to model long-range dependencies and complex stroke patterns. Second, they largely ignore the crucial role of frequency information, which is essential for capturing fine-grained stylistic and structural details in handwriting. To address these challenges, we propose FW-GAN, a one-shot handwriting synthesis framework that generates realistic, writer-consistent text from a single example. Our generator integrates a phase-aware Wave-MLP to better capture spatial relationships while preserving subtle stylistic cues. We further introduce a frequency-guided discriminator that leverages high-frequency components to enhance the authenticity detection of generated samples. Additionally, we introduce a novel Frequency Distribution Loss that aligns the frequency characteristics of synthetic and real handwriting, thereby enhancing visual fidelity. Experiments on Vietnamese and English handwriting datasets demonstrate that FW-GAN generates high-quality, style-consistent handwriting, making it a valuable tool for augmenting data in low-resource handwriting recognition (HTR) pipelines. Official implementation is available at https://github.com/DAIR-Group/FW-GAN

Figures

Figures reproduced from arXiv: 2508.21040 by Dang Hoai Nam, Huynh Tong Dang Khoa, Vo Nguyen Le Duy.

Figure 1
Figure 1. Figure 1: Overview of the proposed FW-GAN generate a corresponding handwritten image xˆk in the style of x, we train a generator G that maps a character sequence and a style vector to an image: xˆk = G(ak, z). (1) There are two ways of obtaining the style vector z: it can either be sampled from a standard normal distribution N (0, I) or extracted directly from the reference image x using a style encoder E by z = E(x… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed WaveGBlock features, enabling personalized and flexible synthesis. To enhance realism and preserve structural coherence, we deploy a dual discriminator architec￾ture where a standard spatial discriminator and a frequency-guided discrim￾inator work in parallel, with the latter leveraging high-frequency wavelet components to detect subtle artifacts that spatial analysis might miss. T… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the Dual Discriminator Architecture. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualization of the transformed frequency components of handwriting images [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of handwriting images generated by each architectural component, [PITH_FULL_IMAGE:figures/full_fig_p026_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of handwriting generated from the IAM dataset across [PITH_FULL_IMAGE:figures/full_fig_p028_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of handwriting generated from the HANDS-VNOnDB [PITH_FULL_IMAGE:figures/full_fig_p028_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative reconstruction results. Each row corresponds to a different model, [PITH_FULL_IMAGE:figures/full_fig_p029_8.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

52 extracted references · 11 canonical work pages · 1 internal anchor

  1. [1]

    H. T. Nguyen, C. T. Nguyen, P. T. Bao, M. Nakagawa, A database of unconstrained vietnamese online handwriting and recognition exper- iments by recurrent neural networks, Pattern Recognition 78 (2018) 291–306. doi:https://doi.org/10.1016/j.patcog.2018.01.013. URL https://www.sciencedirect.com/science/article/pii/ S0031320318300141 30

    doi:10.1016/j.patcog.2018.01.013 2018

  2. [2]

    Kleber, S

    F. Kleber, S. Fiel, M. Diem, R. Sablatnig, CVL-DataBase: An Off-Line Database for Writer Retrieval, Writer Identification and Word Spotting, in: ICDAR, 2013

    2013

  3. [3]

    Pratikakis, K

    I. Pratikakis, K. Zagori, P. Kaddas, B. Gatos, Icfhr 2018 competition on handwritten document image binarization (h-dibco 2018), in: 2018 16th International Conference on Frontiers in Handwriting Recognition (ICFHR), 2018, pp. 489–493.doi:10.1109/ICFHR-2018.2018.00091

    arXiv 2018

  4. [4]

    R. D. Lins, Nabuco - two decades of document processing in latin amer- ica, J. Univers. Comput. Sci. 17 (2011) 151–161. URL https://api.semanticscholar.org/CorpusID:2896293

    work page 2011

  5. [5]

    Graves, Generating Sequences with Recurrent Neural Networks, arXiv preprint arXiv:1308.0850 (2013)

    A. Graves, Generating Sequences with Recurrent Neural Networks, arXiv preprint arXiv:1308.0850 (2013)

    Pith/arXiv arXiv 2013

  6. [6]

    A. K. Bhunia, S. Khan, H. Cholakkal, R. M. Anwer, F. S. Khan, M. Shah, Handwriting Transformers, in: ICCV, 2021

    2021

  7. [7]

    Z.-Q. J. X. Zhi-Qin John Xu, Y. Z. Yaoyu Zhang, T. L. Tao Luo, Y. X. Yanyang Xiao, Z. M. Zheng Ma, Frequency principle: Fourier analysis sheds light on deep neural networks, Communications in Computational Physics 28 (5) (2020) 1746–1767.doi:10.4208/cicp.oa-2020-0085. URL http://dx.doi.org/10.4208/cicp.OA-2020-0085

    doi:10.4208/cicp.oa-2020-0085 2020

  8. [8]

    Y. Tang, K. Han, J. Guo, C. Xu, Y. Li, C. Xu, Y. Wang, An image patch is a wave: Phase-aware vision mlp, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2022, p. 10925–10934. doi:10.1109/cvpr52688.2022.01066. URL http://dx.doi.org/10.1109/CVPR52688.2022.01066

    arXiv 2022

  9. [9]

    Z. Ni, J. Wu, Z. Wang, W. Yang, H. Wang, L. Ma, Misalignment- robust frequency distribution loss for image transformation, in: 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, 2024, p. 2910–2919. doi:10.1109/cvpr52733.2024. 00281. URL http://dx.doi.org/10.1109/CVPR52733.2024.00281

    arXiv 2024

  10. [10]

    Aksan, F

    E. Aksan, F. Pece, O. Hilliges, DeepWriting: Making digital ink editable via deep generative modeling, in: CHI, ACM, 2018. 31

    work page 2018

  11. [11]

    Aksan, O

    E. Aksan, O. Hilliges, STCN: Stochastic Temporal Convolutional Net- works, in: ICLR, 2018

    work page 2018

  12. [12]

    Kotani, S

    A. Kotani, S. Tellex, J. Tompkin, Generating Handwriting via Decou- pled Style Descriptors, in: ECCV, 2020

    2020

  13. [13]

    B. Ji, T. Chen, Generative Adversarial Network for Handwritten Text, arXiv preprint arXiv:1907.11845 (2019)

    Pith/arXiv arXiv 1907

  14. [14]

    J. Wang, C. Wu, Y.-Q. Xu, H.-Y. Shum, Combining Shape and Physical Models for On-line Cursive Handwriting Synthesis, IJDAR 7 (4) (2005) 219–227

    2005

  15. [15]

    Z. Lin, L. Wan, Style-preserving english handwriting synthesis, Pattern Recognit. 40 (7) (2007) 2097–2109

    2007

  16. [16]

    A. O. Thomas, A. Rusu, V. Govindaraju, Synthetic Handwritten CAPTCHAs, Pattern Recognit. 42 (12) (2009) 3365–3373

    2009

  17. [17]

    Haines, O

    T. Haines, O. Mac Aodha, G. Brostow, My Text in Your Handwriting, ACM Trans. Graphics 35 (3) (2016)

    2016

  18. [18]

    Alonso, B

    E. Alonso, B. Moysset, R. Messina, Adversarial Generation of Handwrit- tenTextImagesConditionedonSequences, in: ICDAR,IEEEComputer Society, 2019

    work page 2019

  19. [19]

    Fogel, H

    S. Fogel, H. Averbuch-Elor, S. Cohen, S. Mazor, R. Litman, Scrabble- GAN: Semi-Supervised Varying Length Handwritten Text Generation, in: CVPR, 2020

    work page 2020

  20. [20]

    L. Kang, P. Riba, Y. Wang, M. Rusi˜ nol, A. Fornés, M. Villegas, GAN- writing: Content-Conditioned Generation of Styled Handwritten Word Images, in: ECCV, 2020

    2020

  21. [21]

    J. Gan, W. Wang, HiGAN: Handwriting Imitation Conditioned on Arbitrary-Length Texts and Disentangled Styles, in: AAAI, 2021

    2021

  22. [22]

    Mattick, M

    A. Mattick, M. Mayr, M. Seuret, A. Maier, V. Christlein, SmartPatch: Improving Handwritten Word Imitation with Patch Discriminators, in: ICDAR, 2021. 32

    work page 2021

  23. [23]

    C. Luo, Y. Zhu, L. Jin, Z. Li, D. Peng, SLOGAN: Handwriting Style Synthesis for Arbitrary-Length and Out-of-Vocabulary Text, IEEE Trans. Neural Netw. Learn. Syst. (2022)

    2022

  24. [24]

    Krishnan, R

    P. Krishnan, R. Kovvuri, G. Pang, B. Vassilev, T. Hassner, TextStyle- Brush: Transfer of Text Aesthetics from a Single Example, arXiv e- prints (2021) arXiv–2106

    work page 2021

  25. [25]

    Davis, C

    B. Davis, C. Tensmeyer, B. Price, C. Wigington, B. Morse, R. Jain, Text and Style Conditioned GAN for Generation of Offline Handwriting Lines, in: BMVC, 2020

    2020

  26. [26]

    Brock, J

    A. Brock, J. Donahue, K. Simonyan, Large Scale GAN Training for High Fidelity Natural Image Synthesis, in: ICLR, 2019

    2019

  27. [27]

    Pippi, S

    V. Pippi, S. Cascianelli, R. Cucchiara, Handwritten Text Generation from Visual Archetypes, in: CVPR, 2023

    2023

  28. [28]

    Tolstikhin, N

    I. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Un- terthiner, J. Yung, A. Steiner, D. Keysers, J. Uszkoreit, M. Lucic, A. Dosovitskiy, Mlp-mixer: An all-mlp architecture for vision (2021). arXiv:2105.01601

    Pith/arXiv arXiv 2021

  29. [29]

    X. Ding, C. Xia, X. Zhang, X. Chu, J. Han, G. Ding, Repmlp: Re- parameterizing convolutions into fully-connected layers for image recog- nition (2021). arXiv:2105.01883

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  30. [30]

    Guo, Z.-N

    M.-H. Guo, Z.-N. Liu, T.-J. Mu, S.-M. Hu, Beyond self-attention: Ex- ternal attention using two linear layers for visual tasks, IEEE Trans- actions on Pattern Analysis and Machine Intelligence (2022) 1–13doi: 10.1109/tpami.2022.3211006. URL http://dx.doi.org/10.1109/TPAMI.2022.3211006

    work page arXiv 2022

  31. [31]

    Melas-Kyriazi, Do you even need attention? a stack of feed-forward layers does surprisingly well on imagenet (2021).arXiv:2105.02723

    L. Melas-Kyriazi, Do you even need attention? a stack of feed-forward layers does surprisingly well on imagenet (2021).arXiv:2105.02723

    Pith/arXiv arXiv 2021

  32. [32]

    Z. Liu, Y. Lin, Y. Cao, H. Hu, Y. Wei, Z. Zhang, S. Lin, B. Guo, Swin transformer: Hierarchical vision transformer using shifted windows (2021). arXiv:2103.14030. 33

    Pith/arXiv arXiv 2021

  33. [33]

    J. Ho, N. Kalchbrenner, D. Weissenborn, T. Salimans, Axial attention in multidimensional transformers (2019).arXiv:1912.12180

    Pith/arXiv arXiv 2019

  34. [34]

    T. Luo, Z. Ma, Z.-Q. J. Xu, Y. Zhang, Theory of the frequency princi- ple for general deep neural networks, arXiv preprint arXiv:1906.09235 (2019)

    Pith/arXiv arXiv 1906

  35. [35]

    Jacot, F

    A. Jacot, F. Gabriel, C. Hongler, Neural tangent kernel: Convergence and generalization in neural networks, in: Advances in neural informa- tion processing systems, 2018, pp. 8571–8580

    2018

  36. [36]

    T. Luo, Z. Ma, Z.-Q. J. Xu, Y. Zhang, On the exact computation of linear frequency principle dynamics and its generalization, SIAM Jour- nal on Mathematics of Data Science 4 (4) (2022) 1272–1292. doi: 10.1137/21m1444400. URL http://dx.doi.org/10.1137/21m1444400

    doi:10.1137/21m1444400 2022

  37. [37]

    Zhang, Z.-Q

    Y. Zhang, Z.-Q. J. Xu, T. Luo, Z. Ma, Explicitizing an implicit bias of the frequency principle in two-layer neural networks (2019).arXiv: 1905.10264

    Pith/arXiv arXiv 2019

  38. [38]

    Bordelon, A

    B. Bordelon, A. Canatar, C. Pehlevan, Spectrum dependent learning curves in kernel regression and wide neural networks (2020). arXiv: 2002.02561

    Pith/arXiv arXiv 2020

  39. [39]

    Y. Cao, Z. Fang, Y. Wu, D.-X. Zhou, Q. Gu, Towards understanding the spectral bias of deep learning, arXiv preprint arXiv:1912.01198 (2019)

    Pith/arXiv arXiv 1912

  40. [40]

    Basri, D

    R. Basri, D. Jacobs, Y. Kasten, S. Kritchman, The convergence rate of neural networks for learned functions of different frequencies (2019). arXiv:1906.00425

    Pith/arXiv arXiv 2019

  41. [41]

    G. Yang, H. Salman, A fine-grained spectral perspective on neural net- works, arXiv preprint arXiv:1907.10599 (2019)

    Pith/arXiv arXiv 1907

  42. [42]

    W. E, C. Ma, L. Wu, Machine learning from a continuous viewpoint, arXiv preprint arXiv:1912.12777 (2019)

    Pith/arXiv arXiv 1912

  43. [43]

    J. Gan, W. Wang, J. Leng, X. Gao, HiGAN+: Handwriting Imitation GAN with Disentangled Representations, ACM Trans. Graphics 42 (1) (2022) 1–17. 34

    work page 2022

  44. [44]

    V. B. Braginsky, F. Y. Khalili, K. S. Thorne, Quantum Measurement, Cambridge University Press, 1992

    work page 1992

  45. [45]

    Jacobs, D

    K. Jacobs, D. A. Steck, A straightforward introduction to continuous quantum measurement, Contemporary Physics 47 (5) (2006) 279–303. doi:10.1080/00107510601101934. URL http://dx.doi.org/10.1080/00107510601101934

    doi:10.1080/00107510601101934 2006

  46. [46]

    J. H. Lim, J. C. Ye, Geometric GAN, arXiv preprint arXiv:1705.02894 (2017)

    Pith/arXiv arXiv 2017

  47. [47]

    Graves, S

    A. Graves, S. Fernández, F. Gomez, J. Schmidhuber, Connectionist tem- poral classification: labelling unsegmented sequence data with recurrent neural networks, in: Proceedings of the 23rd International Conference on Machine Learning, ICML ’06, Association for Computing Machin- ery, New York, NY, USA, 2006, p. 369–376. doi:10.1145/1143844. 1143891. URL htt...

    arXiv 2006

  48. [48]

    X. Chen, Y. Duan, R. Houthooft, J. Schulman, I. Sutskever, P. Abbeel, Infogan: Interpretable representation learning by information maximiz- ing generative adversarial nets (2016).arXiv:1606.03657

    Pith/arXiv arXiv 2016

  49. [49]

    J.-Y. Zhu, R. Zhang, D. Pathak, T. Darrell, A. A. Efros, O. Wang, E. Shechtman, Toward multimodal image-to-image translation (2017). arXiv:1711.11586

    Pith/arXiv arXiv 2017

  50. [50]

    Lee, H.-Y

    H.-Y. Lee, H.-Y. Tseng, J.-B. Huang, M. Singh, M.-H. Yang, Di- verse Image-to-Image Translation via Disentangled Representations, Springer International Publishing, 2018, p. 36–52. doi:10.1007/ 978-3-030-01246-5_3. URL http://dx.doi.org/10.1007/978-3-030-01246-5_3

    doi:10.1007/978-3-030-01246-5_3 2018

  51. [51]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, in: ICLR, 2015

    2015

  52. [52]

    M. Li, T. Lv, J. Chen, L. Cui, Y. Lu, D. Florencio, C. Zhang, Z. Li, F. Wei, Trocr: Transformer-based optical character recognition with pre-trained models, Proceedings of the AAAI Conference on Artificial Intelligence 37 (11) (2023) 13094–13102. doi:10.1609/aaai.v37i11. 35 26538. URL http://dx.doi.org/10.1609/aaai.v37i11.26538 36

    doi:10.1609/aaai.v37i11 2023

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.