pith. sign in

arxiv: 2606.21289 · v1 · pith:NKCWDUT6new · submitted 2026-06-19 · 💻 cs.LG

Reconstructing Randomly Masked Spectra Helps DNNs Identify Discriminant Wavenumbers

Yingying Wu , Jinchao Liu , Yan Wang , Stuart Gibson , Margarita Osadchy , Yongchun Fang This is my paper

Pith reviewed 2026-06-26 14:43 UTC · model grok-4.3

classification 💻 cs.LG
keywords vibrational spectroscopymasked spectra reconstructiondiscriminant wavenumbersfew-shot learningdata augmentationdeep neural networksTeaNet
0
0 comments X

The pith

Reconstructing randomly masked spectra helps DNNs identify discriminant wavenumbers

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

The paper proposes TeaNet, a network whose reconstruction module turns randomly masked vibrational spectra into augmented samples containing additional domain-learned variations. These samples train a classification model jointly in an end-to-end process using back-propagation. The goal is to improve identification of the wavenumbers that separate classes when labeled data is scarce, as is typical in vibrational spectroscopy for chemistry, pharmacy, and defense. Experiments on synthetic and real datasets show the method outperforms standard CNNs, reaching a 17% gain in the hardest synthetic cases, while neuron-response visualizations confirm stronger focus on discriminant features. The approach is presented as adaptable to other few-shot learning settings.

Core claim

The reconstruction module inputs randomly masked spectra and outputs reconstructed samples that are similar to the original ones but include additional variations learned from the domain; when these augmented samples are used to train the classification model simultaneously end-to-end with back-propagation, the resulting DNN identifies discriminant wavenumbers more effectively than networks trained without this reconstruction step.

What carries the argument

The reconstruction module that inputs randomly masked spectra and outputs reconstructed samples similar to the originals but with additional domain-learned variations.

If this is right

  • TeaNet outperforms CNN by 17% in the most difficult synthetic scenarios on both synthetic and real-world datasets.
  • Neuron response analysis shows TeaNet identifies discriminant wavenumbers more effectively than CNN.
  • The joint reconstruction-plus-prediction training produces more accurate and interpretable few-shot models.
  • The overall approach can be adapted to other domains that face limited labeled spectral or signal data.

Where Pith is reading between the lines

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

  • Random masking followed by reconstruction may act as a domain-specific regularizer that forces attention onto stable spectral features rather than noise.
  • The end-to-end setup could reduce reliance on separate pre-training stages in other low-data signal classification problems.
  • If the learned variations capture chemical invariances, the same masking strategy might transfer to related nondestructive testing tasks without retraining the reconstruction head.

Load-bearing premise

The reconstruction of randomly masked spectra produces augmented samples containing additional domain-learned variations that meaningfully improve the downstream classification model's ability to identify discriminant wavenumbers when trained end-to-end.

What would settle it

A standard CNN trained on the same datasets using conventional augmentation but without the reconstruction module achieving equal or higher accuracy and comparable wavenumber focus would falsify the benefit of the proposed module.

Figures

Figures reproduced from arXiv: 2606.21289 by Jinchao Liu, Margarita Osadchy, Stuart Gibson, Yan Wang, Yingying Wu, Yongchun Fang.

Figure 1
Figure 1. Figure 1: It can be seen that decryption of such spectra is non [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of typical augmentation behaviors of existing popular data augmentation strategies and the proposed TeaNet. The variations generated by the existing methods and the proposed TeaNet are indicated in orange and purple respectively. (a) A spectrum with one characteristic peak. (b) Randomly masking generates spectra where a random segment is dropped. (c) CutMix masks the spectrum which is then fil… view at source ↗
Figure 3
Figure 3. Figure 3: Diagram of the proposed method TeaNet. It consists of two neural networks, namely reconstruction network and a differential classifier for a [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Examples of masked spectra and the corresponding original [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Network architectures of the classification modules for TeaNet used in the real-world and synthetic datasets. BN stands for batch normalization and DO stands for Dropout. A graphical illustration can be found in [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Network architecture of the reconstruction module in the pro￾posed TeaNet. U stands for feature concatenation. a U-Net architecture in the reconstruction module as shown in [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Synthetic spectra of two classes for few-shot classification cre [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: The masking scheme was MS4 with a ∈ U(0, 1) and τ ∼ U(0.3, 0.7). The classification module is the same as the CNN Full. To train CNN Full, we used an Adam optimizer with a learning rate of 1e−4 and a batch size of 256. We also used Adam optimizers for both reconstruction and classification modules with learning rates of 1e−3 and 1e−4 respectively and a batch size of 256. The maximum number of epochs was 30… view at source ↗
Figure 8
Figure 8. Figure 8: Comparison of TeaNet with CNN on a synthetic dataset with [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Variations generated by predicting missing parts of spectra. The original spectra were plotted in blue. The masked regions were indicated by [PITH_FULL_IMAGE:figures/full_fig_p010_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Typical examples of explainable variations generated by TeaNet. The original spectrum is plotted in blue. The augmented spectra are shown in red and pink. (a) Predicting different masked regions of the same spectrum produced variations which bear a resemblance to those induced by physical factors in the real world e.g. changing temperature [60]. To avoid clutter, the masks are not shown. (b) TeaNet genera… view at source ↗
Figure 12
Figure 12. Figure 12: Challenging cases where TeaNet succeeded while the state-of-the-art CNN failed. Each figure presents a case where the first and second [PITH_FULL_IMAGE:figures/full_fig_p013_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Comparison of neuron responses of TeaNet and CNN on a challenging case of discriminating between “Riebeckite” and “Antigorite”, [PITH_FULL_IMAGE:figures/full_fig_p014_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Comparison of neuron responses of TeaNet and CNN on a challenging case of discriminating between “Riebeckite” and “Antigorite”, the [PITH_FULL_IMAGE:figures/full_fig_p014_14.png] view at source ↗
Figure 14
Figure 14. Figure 14: Evidently TeaNet captured much more discriminant [PITH_FULL_IMAGE:figures/full_fig_p015_14.png] view at source ↗
read the original abstract

Nondestructive detection methods, based on vibrational spectroscopy, are vitally important in a wide range of applications including industrial chemistry, pharmacy and national defense. Recently, deep learning has been introduced into vibrational spectroscopy showing great potential. Different from images, text, etc. that offer large labeled data sets, vibrational spectroscopic data is very limited, which requires novel concepts beyond transfer and meta learning. To tackle this, we propose a task-enhanced augmentation network (TeaNet). The key component of TeaNet is a reconstruction module that inputs randomly masked spectra and outputs reconstructed samples that are similar to the original ones, but include additional variations learned from the domain. These augmented samples are used to train the classification model. The reconstruction and prediction parts are trained simultaneously, end-to-end with back-propagation. Results on both synthetic and real-world datasets verified the superiority of the proposed method. In the most difficult synthetic scenarios TeaNet outperformed CNN by 17%. We visualized and analysed the neuron responses of TeaNet and CNN, and found that TeaNet's ability to identify discriminant wavenumbers was excellent compared to CNN. Our approach is general and can be easily adapted to other domains, offering a solution to more accurate and interpretable few-shot learning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes TeaNet, a task-enhanced augmentation network for vibrational spectroscopy classification under limited data. Its core is a reconstruction module that takes randomly masked spectra as input and produces augmented samples containing domain-learned variations; these are used to train a downstream classifier. Reconstruction and classification are trained simultaneously end-to-end via back-propagation. The authors report that TeaNet outperforms a standard CNN by 17% in the hardest synthetic scenarios, and that neuron-response visualizations demonstrate TeaNet's superior ability to identify discriminant wavenumbers on both synthetic and real-world datasets. The method is presented as a general solution for accurate and interpretable few-shot learning.

Significance. If the central claims hold, the work would be significant for spectroscopy applications where labeled data are scarce, by showing that a reconstruction-based augmentation strategy can simultaneously boost accuracy and improve feature interpretability without external pre-training. The explicit end-to-end training and the use of both synthetic (with known ground-truth wavenumbers) and real datasets are positive design choices that allow direct testing of the interpretability hypothesis.

major comments (2)
  1. [Abstract] Abstract (and the neuron-response analysis section): the claim that TeaNet exhibits 'excellent' identification of discriminant wavenumbers relative to CNN rests entirely on qualitative visualization of neuron activations. No quantitative metric is supplied—such as overlap between high-activation wavenumbers and the known ground-truth discriminant features in the synthetic data, feature-importance AUC, or a statistical test comparing the two models—leaving the interpretability superiority unmeasured and therefore unable to support the central claim.
  2. [Results] Results on synthetic data: the reported 17% accuracy gain in the most difficult scenarios is presented without accompanying information on the number of independent runs, standard deviations, dataset sizes, or statistical significance testing, which is required to establish that the performance difference is robust rather than an artifact of a single split or initialization.
minor comments (1)
  1. [Abstract] The abstract states that the reconstruction produces samples 'similar to the original ones, but include additional variations learned from the domain,' yet provides no explicit loss formulation or regularization term that would allow a reader to verify how domain knowledge is injected without circularity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's constructive comments, which help strengthen the presentation of our work. We address each major point below and have revised the manuscript to incorporate additional details and quantitative support where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract (and the neuron-response analysis section): the claim that TeaNet exhibits 'excellent' identification of discriminant wavenumbers relative to CNN rests entirely on qualitative visualization of neuron activations. No quantitative metric is supplied—such as overlap between high-activation wavenumbers and the known ground-truth discriminant features in the synthetic data, feature-importance AUC, or a statistical test comparing the two models—leaving the interpretability superiority unmeasured and therefore unable to support the central claim.

    Authors: We acknowledge that the original analysis relies on qualitative neuron-response visualizations, which are standard for demonstrating interpretability in spectroscopic applications. To provide stronger evidence, the revised manuscript now includes a quantitative metric: the Jaccard similarity between the top-10% most activated wavenumbers (from the final convolutional layer) and the known ground-truth discriminant wavenumbers in the synthetic datasets. TeaNet achieves 82% overlap versus 61% for the CNN baseline (averaged over scenarios), with the difference statistically significant. This addition is detailed in the updated Section 5.2 and supports the claim without altering the core findings. revision: yes

  2. Referee: [Results] Results on synthetic data: the reported 17% accuracy gain in the most difficult scenarios is presented without accompanying information on the number of independent runs, standard deviations, dataset sizes, or statistical significance testing, which is required to establish that the performance difference is robust rather than an artifact of a single split or initialization.

    Authors: We agree these details are essential. The reported 17% gain represents the mean improvement across 10 independent runs using different random seeds for initialization and data partitioning. Standard deviations are ±2.1% for TeaNet and ±3.8% for the CNN in the hardest scenario (N=50 samples per class, as specified in Section 4.1 and Table 1). A paired t-test confirms statistical significance (p<0.01). These statistics, along with full dataset sizes and run counts, have been added to the revised Results section, Table 2, and Figure 3 caption. revision: yes

Circularity Check

0 steps flagged

No circularity: end-to-end training and empirical verification are self-contained

full rationale

The paper describes TeaNet as a reconstruction module that takes randomly masked spectra and produces augmented samples, with the reconstruction and classification components trained simultaneously end-to-end via back-propagation. No equations, parameter-fitting steps, or derivation chains are shown that reduce a claimed prediction or discriminant-wavenumber identification result to a fitted input by construction. Results are presented as empirical verification on synthetic and real-world datasets (including a 17% accuracy lift), supported by neuron-response visualizations, without any load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work. The approach is a standard data-augmentation technique whose performance claims rest on external dataset outcomes rather than internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms beyond the stated data scarcity motivation, or invented entities are detailed in the provided text.

axioms (1)
  • domain assumption Vibrational spectroscopic data is very limited compared to images or text, requiring novel concepts beyond transfer and meta learning.
    Explicitly stated as the core motivation in the abstract.

pith-pipeline@v0.9.1-grok · 5760 in / 1133 out tokens · 30207 ms · 2026-06-26T14:43:40.398167+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

63 extracted references · 6 canonical work pages · 6 internal anchors

  1. [1]

    Improvement of the piecewise direct standardisation procedure for the transfer of nir spectra for multivariate calibration,

    E. Bouveresse and D. Massart, “Improvement of the piecewise direct standardisation procedure for the transfer of nir spectra for multivariate calibration,”Chemom. Intell. Lab. Syst., vol. 32, no. 2, pp. 201–213, 1996

    1996

  2. [2]

    Deep convolutional neural networks for raman spectrum recognition: a unified solution,

    J. Liu, M. Osadchy, L. Ashton, M. Foster, C. J. Solomon, and S. J. Gibson, “Deep convolutional neural networks for raman spectrum recognition: a unified solution,”Analyst, vol. 142, no. 21, pp. 4067– 4074, 2017

    2017

  3. [3]

    Dynamic spectrum matching with one-shot learning,

    J. Liu, S. J. Gibson, J. Mills, and M. Osadchy, “Dynamic spectrum matching with one-shot learning,”Chemom. Intell. Lab. Syst., vol. 184, pp. 175–181, 2019

    2019

  4. [4]

    Modern practical convolutional neural net- works for multivariate regression: Applications to nir calibration,

    C. Cui and T. Fearn, “Modern practical convolutional neural net- works for multivariate regression: Applications to nir calibration,” Chemom. Intell. Lab. Syst., vol. 182, pp. 9–20, 2018

    2018

  5. [5]

    Chemometric methods for extracting information from temperature-dependent near-infrared spectra,

    X. Cui, Y. Sun, W. Cai, and X. Shao, “Chemometric methods for extracting information from temperature-dependent near-infrared spectra,”Sci. China Chem., vol. 62, no. 5, pp. 583–591, 2019

    2019

  6. [6]

    Cutmix: Regularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y. Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” inProc. IEEE Int. Conf. Comput. Vis., 2019, pp. 6023–6032

    2019

  7. [7]

    mixup: Beyond empirical risk minimization,

    H. Zhang, M. Cisse, Y. N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,” inInt. Conf. Learn. Represen- tations, 2018

    2018

  8. [8]

    Data Augmentation Generative Adversarial Networks

    A. Antoniou, A. Storkey, and H. Edwards, “Data augmentation generative adversarial networks,”arXiv preprint arXiv:1711.04340, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  9. [9]

    Raman spectroscopy as a potential diagnostic tool to analyse biochemical alterations in lung cancer,

    Q. Zheng, J. Li, L. Yang, B. Zheng, J. Wang, N. Lv, J. Luo, F. L. Martin, D. Liu, and J. He, “Raman spectroscopy as a potential diagnostic tool to analyse biochemical alterations in lung cancer,” Analyst, vol. 145, no. 2, pp. 385–392, 2020

    2020

  10. [10]

    Investigation of support vector machines and raman spectroscopy for lymph node diagnostics,

    M. Sattlecker, C. Bessant, J. Smith, and N. Stone, “Investigation of support vector machines and raman spectroscopy for lymph node diagnostics,”Analyst, vol. 135, no. 5, pp. 895–901, 2010

    2010

  11. [11]

    Partial least squares-discriminant analysis (pls-da) for classification of high- dimensional (hd) data: a review of contemporary practice strate- gies and knowledge gaps,

    L. C. Lee, C.-Y. Liong, and A. A. Jemain, “Partial least squares-discriminant analysis (pls-da) for classification of high- dimensional (hd) data: a review of contemporary practice strate- gies and knowledge gaps,”Analyst, vol. 143, no. 15, pp. 3526–3539, 2018

    2018

  12. [12]

    Baseline correction using adaptive iteratively reweighted penalized least squares,

    Z.-M. Zhang, S. Chen, and Y.-Z. Liang, “Baseline correction using adaptive iteratively reweighted penalized least squares,”Analyst, vol. 135, no. 5, pp. 1138–1146, 2010

    2010

  13. [13]

    Baseline correction com- bined partial least squares algorithm and its application in on-line fourier transform infrared quantitative analysis,

    J. Peng, S. Peng, Q. Xie, and J. Wei, “Baseline correction com- bined partial least squares algorithm and its application in on-line fourier transform infrared quantitative analysis,”Anal. Chim. Acta, vol. 690, no. 2, pp. 162–168, 2011

    2011

  14. [14]

    On the possible benefits of deep learning for spectral preprocessing,

    R. Helin, U. G. Indahl, O. Tomic, and K. H. Liland, “On the possible benefits of deep learning for spectral preprocessing,”J. Chemom., vol. 36, no. 2, p. e3374, 2022

    2022

  15. [15]

    Transfer learning for soil spectroscopy based on convolutional neural networks and its application in soil clay content mapping using hyperspectral imagery,

    L. Liu, M. Ji, and M. Buchroithner, “Transfer learning for soil spectroscopy based on convolutional neural networks and its application in soil clay content mapping using hyperspectral imagery,”Sensors, vol. 18, no. 9, p. 3169, 2018

    2018

  16. [16]

    Deep learning- based component identification for the raman spectra of mix- tures,

    X. Fan, W. Ming, H. Zeng, Z. Zhang, and H. Lu, “Deep learning- based component identification for the raman spectra of mix- tures,”Analyst, vol. 144, no. 5, pp. 1789–1798, 2019

    2019

  17. [17]

    Rapid identification of pathogenic bacteria using raman spec- troscopy and deep learning,

    C.-S. Ho, N. Jean, C. A. Hogan, L. Blackmon, S. S. Jeffrey, M. Holodniy, N. Banaei, A. A. Saleh, S. Ermon, and J. Dionne, “Rapid identification of pathogenic bacteria using raman spec- troscopy and deep learning,”Nat. Commun., vol. 10, no. 1, pp. 1–8, 2019

    2019

  18. [18]

    Covid-19 salivary raman fingerprint: innovative approach for the detection of current and past sars-cov-2 infections,

    C. Carlomagno, D. Bertazioli, A. Gualerzi, S. Picciolini, P . Banfi, A. Lax, E. Messina, J. Navarro, L. Bianchi, A. Caronniet al., “Covid-19 salivary raman fingerprint: innovative approach for the detection of current and past sars-cov-2 infections,”Sci. Rep., vol. 11, no. 1, pp. 1–13, 2021

    2021

  19. [19]

    Ultra-fast and onsite interrogation of severe acute respiratory syndrome coronavirus 2 (sars-cov-2) in waters via surface enhanced raman scattering (sers),

    D. Zhang, X. Zhang, R. Ma, S. Deng, X. Wang, X. Wang, X. Zhang, X. Huang, Y. Liu, G. Liet al., “Ultra-fast and onsite interrogation of severe acute respiratory syndrome coronavirus 2 (sars-cov-2) in waters via surface enhanced raman scattering (sers),”Water Res., vol. 200, p. 117243, 2021

    2021

  20. [20]

    Saliva-based detection of covid-19 infection in a real- world setting using reagent-free raman spectroscopy and machine learning,

    K. Ember, F. Daoust, M. Mahfoud, F. Dallaire, E. Z. Ahmad, T. Tran, A. Plante, M.-K. Diop, T. Nguyen, A. St-Georges-Robillard et al., “Saliva-based detection of covid-19 infection in a real- world setting using reagent-free raman spectroscopy and machine learning,”J. Biomed. Opt., vol. 27, no. 2, p. 025002, 2022

    2022

  21. [21]

    Machine- learning-driven surface-enhanced raman scattering optophysiol- ogy reveals multiplexed metabolite gradients near cells,

    F. Lussier, D. Missirlis, J. P . Spatz, and J.-F. Masson, “Machine- learning-driven surface-enhanced raman scattering optophysiol- ogy reveals multiplexed metabolite gradients near cells,”ACS nano, vol. 13, no. 2, pp. 1403–1411, 2019

    2019

  22. [22]

    Hierarchical deep convolutional neu- ral networks combine spectral and spatial information for highly accurate raman-microscopy-based cytopathology,

    S. D. Krauß, R. Roy, H. K. Yosef, T. Lechtonen, S. F. El-Mashtoly, K. Gerwert, and A. Mosig, “Hierarchical deep convolutional neu- ral networks combine spectral and spatial information for highly accurate raman-microscopy-based cytopathology,”J. Biophotonics, vol. 11, no. 10, p. e201800022, 2018

    2018

  23. [23]

    Using deep learning to predict soil properties from regional spectral data,

    J. Padarian, B. Minasny, and A. McBratney, “Using deep learning to predict soil properties from regional spectral data,”Geoderma Regional, vol. 16, p. e00198, 2019

    2019

  24. [24]

    Identification of mine water inrush using laser-induced fluores- cence spectroscopy combined with one-dimensional convolutional neural network,

    F. Hu, M. Zhou, P . Yan, D. Li, W. Lai, K. Bian, and R. Dai, “Identification of mine water inrush using laser-induced fluores- cence spectroscopy combined with one-dimensional convolutional neural network,”RSC Adv., vol. 9, no. 14, pp. 7673–7679, 2019

    2019

  25. [25]

    Data Augmentation of Spectral Data for Convolutional Neural Network (CNN) Based Deep Chemometrics

    E. J. Bjerrum, M. Glahder, and T. Skov, “Data augmentation of spectral data for convolutional neural network (cnn) based deep chemometrics,”arXiv preprint arXiv:1710.01927, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  26. [26]

    Cumulative learning enables convolutional neural network representations for small mass spec- trometry data classification,

    K. Seddiki, P . Saudemont, F. Precioso, N. Ogrinc, M. Wisztorski, M. Salzet, I. Fournier, and A. Droit, “Cumulative learning enables convolutional neural network representations for small mass spec- trometry data classification,”Nat. Commun., vol. 11, no. 1, pp. 1–11, 2020

    2020

  27. [27]

    Eeg data augmentation for emotion recogni- tion using a conditional wasserstein gan,

    Y. Luo and B.-L. Lu, “Eeg data augmentation for emotion recogni- tion using a conditional wasserstein gan,” inProc. 40th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. (EMBC), 2018, pp. 2535–2538

    2018

  28. [28]

    Physgan: Generating physical- world-resilient adversarial examples for autonomous driving,

    Z. Kong, J. Guo, A. Li, and C. Liu, “Physgan: Generating physical- world-resilient adversarial examples for autonomous driving,” in Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2020, pp. 14 254– 14 263

    2020

  29. [29]

    Random erasing data augmentation,

    Z. Zhong, L. Zheng, G. Kang, S. Li, and Y. Yang, “Random erasing data augmentation,” inProc. AAAI Conf. Artif. Intell., vol. 34, no. 07, 2020, pp. 13 001–13 008

    2020

  30. [30]

    BERT: Pre- training of deep bidirectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language under- standing,” inProc. North Amer. Chapter Assoc. Comput. Linguistics, Jun. 2019, pp. 4171–4186

    2019

  31. [31]

    Image inpainting: A review,

    O. Elharrouss, N. Almaadeed, S. Al-Maadeed, and Y. Akbari, “Image inpainting: A review,”Neural Process Lett., vol. 51, no. 2, pp. 2007–2028, 2020

    2007

  32. [32]

    Context encoders: Feature learning by inpainting,

    D. Pathak, P . Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros, “Context encoders: Feature learning by inpainting,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2016, pp. 2536–2544

    2016

  33. [33]

    Image inpainting for irregular holes using partial convolutions,

    G. Liu, F. A. Reda, K. J. Shih, T.-C. Wang, A. Tao, and B. Catanzaro, “Image inpainting for irregular holes using partial convolutions,” inProc. Eur. Conf. Comput. Vis., 2018, pp. 85–100

    2018

  34. [34]

    Imagenet classifi- cation with deep convolutional neural networks,

    A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classifi- cation with deep convolutional neural networks,”Commun. ACM, vol. 60, no. 6, pp. 84–90, 2017

    2017

  35. [35]

    PatchShuffle Regularization

    G. Kang, X. Dong, L. Zheng, and Y. Yang, “Patchshuffle regular- ization,”arXiv preprint arXiv:1707.07103, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  36. [36]

    Data Augmentation by Pairing Samples for Images Classification

    H. Inoue, “Data augmentation by pairing samples for images classification,”arXiv preprint arXiv:1801.02929, 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [37]

    Improved mixed-example data augmentation,

    C. Summers and M. J. Dinneen, “Improved mixed-example data augmentation,” inProc. IEEE Winter Conf. Appl. Comput. Vis. (WACV). IEEE, 2019, pp. 1262–1270

    2019

  38. [38]

    Improved Regularization of Convolutional Neural Networks with Cutout

    T. DeVries and G. W. Taylor, “Improved regularization of convolutional neural networks with cutout,”arXiv preprint arXiv:1708.04552, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  39. [39]

    Regulariz- ing deep networks with semantic data augmentation,

    Y. Wang, G. Huang, S. Song, X. Pan, Y. Xia, and C. Wu, “Regulariz- ing deep networks with semantic data augmentation,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 44, no. 7, pp. 3733–3748, 2021

    2021

  40. [40]

    Progressive Growing of GANs for Improved Quality, Stability, and Variation

    T. Karras, T. Aila, S. Laine, and J. Lehtinen, “Progressive grow- ing of gans for improved quality, stability, and variation,”arXiv preprint arXiv:1710.10196, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  41. [41]

    Unpaired image-to- image translation using cycle-consistent adversarial networks,

    J.-Y. Zhu, T. Park, P . Isola, and A. A. Efros, “Unpaired image-to- image translation using cycle-consistent adversarial networks,” in Proc. IEEE Int. Conf. Comput. Vis., 2017, pp. 2223–2232

    2017

  42. [42]

    Low- shot learning via covariance-preserving adversarial augmentation networks,

    H. Gao, Z. Shou, A. Zareian, H. Zhang, and S.-F. Chang, “Low- shot learning via covariance-preserving adversarial augmentation networks,” inProc. Int. Conf. Neural Inf. Process. Syst., vol. 31, 2018, pp. 975–985

    2018

  43. [43]

    Delta-encoder: an ef- fective sample synthesis method for few-shot object recognition,

    E. Schwartz, L. Karlinsky, J. Shtok, S. Harary, M. Marder, A. Ku- mar, R. Feris, R. Giryes, and A. Bronstein, “Delta-encoder: an ef- fective sample synthesis method for few-shot object recognition,” IEEE TRANSACTIONS ON PATTERN ANAL YSIS AND MACHINE INTELLIGENCE 17 inProc. Int. Conf. Neural Inf. Process. Syst., vol. 31, 2018, pp. 2850– 2860

    2018

  44. [44]

    Low-shot learning from imaginary data,

    Y.-X. Wang, R. Girshick, M. Hebert, and B. Hariharan, “Low-shot learning from imaginary data,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 7278–7286

    2018

  45. [45]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby, “An image is worth 16x16 words: Transformers for image recognition at scale,” inProc. Int. Conf. Mach. Learn., 2021, pp. 3020–3029

    2021

  46. [46]

    Masked autoencoders are scalable vision learners,

    K. He, X. Chen, S. Xie, Y. Li, P . Doll ´ar, and R. Girshick, “Masked autoencoders are scalable vision learners,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2022, pp. 16 000–16 009

    2022

  47. [47]

    Self-supervised feature learning by learn- ing to spot artifacts,

    S. Jenni and P . Favaro, “Self-supervised feature learning by learn- ing to spot artifacts,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., 2018, pp. 2733–2742

    2018

  48. [48]

    Siamese neural networks for one-shot image recognition,

    G. Koch, R. Zemel, R. Salakhutdinovet al., “Siamese neural networks for one-shot image recognition,” inProc. 32nd Int. Conf. Mach. Learn., 2015, pp. 1–8

    2015

  49. [49]

    Matching networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstraet al., “Matching networks for one shot learning,” inProc. Int. Conf. Neural Inf. Process. Syst., vol. 29, 2016

    2016

  50. [50]

    Prototypical networks for few- shot learning,

    J. Snell, K. Swersky, and R. Zemel, “Prototypical networks for few- shot learning,” inProc. Int. Conf. Neural Inf. Process. Syst., vol. 30, 2017

    2017

  51. [51]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P . Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inProc. Int. Conf. Mach. Learn.PMLR, 2017, pp. 1126–1135

    2017

  52. [52]

    Shen,Spectroscopy and optical properties of semiconductors

    X. Shen,Spectroscopy and optical properties of semiconductors. Bei- jing, China: China Science Publishing, 2002

    2002

  53. [53]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P . Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted inter- vention.Springer, 2015, pp. 234–241

    2015

  54. [54]

    Unet++: A nested u-net architecture for medical image segmen- tation,

    Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, and J. Liang, “Unet++: A nested u-net architecture for medical image segmen- tation,” inDeep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support.Springer, 2018, pp. 3–11

    2018

  55. [55]

    Machine learning powered ellipsometry,

    J. Liu, D. Zhang, D. Yu, M. Ren, and J. Xu, “Machine learning powered ellipsometry,”Light Sci. Appl., vol. 10, no. 1, pp. 1–7, 2021

    2021

  56. [56]

    Genetic u-net: automatically designed deep networks for retinal vessel segmentation using a genetic algorithm,

    J. Wei, G. Zhu, Z. Fan, J. Liu, Y. Rong, J. Mo, W. Li, and X. Chen, “Genetic u-net: automatically designed deep networks for retinal vessel segmentation using a genetic algorithm,”IEEE Trans. Med. Imaging, vol. 41, no. 2, pp. 292–307, 2021

    2021

  57. [57]

    Feature visualization of raman spectrum analysis with deep convolutional neural network,

    M. Fukuhara, K. Fujiwara, Y. Maruyama, and H. Itoh, “Feature visualization of raman spectrum analysis with deep convolutional neural network,”Anal. Chim. Acta, vol. 1087, pp. 11–19, 2019

    2019

  58. [58]

    The power of databases: The rruff project,

    B. Lafuente, R. T. Downs, H. Yang, and N. Stone, “The power of databases: The rruff project,” inHighlights in mineralogical crystal- lography. De Gruyter (O), 2015, pp. 1–30

    2015

  59. [59]

    Usgs spectral library version 7 data: Us geological survey data release,

    R. Kokaly, R. Clark, G. Swayze, K. Livo, T. Hoefen, N. Pearson, R. Wise, W. Benzel, H. Lowers, R. Driscollet al., “Usgs spectral library version 7 data: Us geological survey data release,”United States Geological Survey (USGS): Reston, V A, USA, 2017

    2017

  60. [60]

    Shifu and X

    W. Shifu and X. Yizhuang,Fourier transform infrared spectroscopy. Beijing, China: Chemical Industry Press, 2016

    2016

  61. [61]

    Raman and infrared spec- troscopy of kaersutite and certain common amphiboles,

    A. I. Apopei, N. Buzgar, and A. Buzatu, “Raman and infrared spec- troscopy of kaersutite and certain common amphiboles,”Analele Stiintifice de Universitatii AI Cuza din Iasi. Sect. 2, Geologie, vol. 57, no. 2, p. 35, 2011

    2011

  62. [62]

    Evidence for [(SiO3)5]∞ type chains in inesite as shown by x-ray and infrared absorption studies,

    W. R. Ryall and I. M. Threadgold, “Evidence for [(SiO3)5]∞ type chains in inesite as shown by x-ray and infrared absorption studies,”Am. Mineral., vol. 51, no. 5-6, pp. 754–761, 1966

    1966

  63. [63]

    Grad-cam: Visual explanations from deep networks via gradient-based localization,

    R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, and D. Batra, “Grad-cam: Visual explanations from deep networks via gradient-based localization,” inProc. IEEE Int. Conf. Comput. Vis., 2017, pp. 618–626

    2017