Pith. sign in

REVIEW 2 major objections 5 minor 41 references

Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation

T0 review · 2 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Masked signal pretraining on unlabeled RF data gives a 17.5% one-shot radar accuracy improvement over training from scratch.

desk verdict Useful benchmark paper for few-shot radar recognition, but the headline SSL gains are likely upper bounds because the masking hyperparameters appear to be selected on the test set. read the letter →

arxiv 2501.03461 v3 pith:243XX6PG submitted 2025-01-07 cs.LG cs.AIeess.SP

Zi Huang , Simon Denman , Akila Pemasiri , Clinton Fookes , Terrence Martin This is my paper · ORCID
classification cs.LGcs.AIeess.SP
keywords few-shotlearningself-supervisedmaskedsignalmodellingautoencodersradarrecognitiondomainadaptationI/QsignalsRFclassification
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 sets out to establish that self-supervised pre-training by masked signal modelling on unlabeled I/Q radio data can substitute for large amounts of labelled radar data. After reconstructing corrupted baseband signals, a lightweight encoder fine-tuned on just one labelled frame per radar class and SNR level outperforms the same model trained from scratch: ResNet1D gains 17.5% relative accuracy when pre-trained on in-domain radar signals and 16.31% when pre-trained on out-of-domain communications signals. A sympathetic reader would care because electronic-warfare settings often have few radar labels and require rapid model re-training, so any representation that transfers from unlabeled RF collections is practically valuable. The paper also provides the first benchmark comparing masked autoencoder designs, masking strategies, and pre-training domains for few-shot radar classification.

What carries the argument

The load-bearing mechanism is masked signal modelling (MSM), applied to baseband in-phase/quadrature (I/Q) sequences. The paper defines four masking strategies—random zero-masking (strategy A), random block zero-masking (strategy B), random noise-masking (strategy C), and block noise-masking (strategy D)—with a masking ratio $R_m$; the masked sequence is passed through an asymmetric masked autoencoder trained with an $\ell_1$ sample-wise reconstruction loss, and $R_m$ governs how much of the signal is obscured. After pre-training, the decoder is discarded, a linear probing classifier is attached to the encoder, and only the classifier is fine-tuned on a few annotated frames. This two-step recipe carries the few-shot transfer, and the paper's comparison of $S_m$ and $R_m$ across four source RF datasets establishes that the optimal masking choices are domain-dependent.

What would settle it

Re-run the 1-shot and 5-shot experiments for ResNet1D and WaveNet with the masking strategy and ratio fixed in advance (or tuned on a held-out split of the pre-training domain), then compare against the numbers in Table 2; if the gaps over the no-SSL baselines vanish or reverse, the central claim is not robust.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that masked signal modelling is an effective annotation-free pre-training task for few-shot radar signal recognition, and that pre-training on diverse RF domains (radar, communications, or a mixture) transfers to the radar domain when the target has only a few labelled frames. The authors demonstrate this by pre-training ResNet1D, MS-TCN, and WaveNet autoencoders to reconstruct corrupted I/Q signals, then replacing the decoder with a linear probing classifier and fine-tuning on RadChar-nShot (205 frames in the 1-shot case, 2050 in the 10-shot case) before evaluating on RadChar-Eval. In their tables, the best masking strategy and ratio vary by model and source domain, but every model pre-trained on in-domain RadChar-SSL improves over its no-SSL baseline in the 1-shot setting, and ResNet1D's largest gain is a 17.5% accuracy increase; pre-training on RadioML, an out-of-domain communications dataset, still yields a 16.31% gain for ResNet1D.

Load-bearing premise

The headline gains rest on the assumption that the per-domain optimal masking strategy and ratio were selected without using the test set, yet Section 3.1 reports that no validation split is used during fine-tuning, so if those settings were chosen on RadChar-Eval itself, the reported improvements are optimistic upper bounds.

Editorial extensions

If this is right

  • If the central claim is correct, an unlabeled corpus of I/Q signals from any accessible RF domain can serve as pre-training data for a radar classifier that will be fine-tuned with only a few labelled pulses.
  • The reported saturation at roughly 10 shots suggests that the practical benefit of SSL is concentrated in the most label-starved regime, after which additional labelled frames matter more than pre-training.
  • The optimal masking strategy and ratio are not universal: the same ResNet1D prefers low-ratio random zero-masking on RadioML and DeepRadar but high-ratio random zero-masking on RadChar-SSL, so practitioners should tune these choices per source domain.
  • Gains are concentrated at moderate-to-high signal-to-noise ratios, so SSL pre-training will help most when the intercepted signal is not buried in noise.

Reading between the lines

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

  • An extension the paper leaves implicit is that the same masked-signal recipe could apply to other RF label-scarce tasks such as specific emitter identification, where only a handful of captures per emitter exist; the out-of-domain transfer result suggests the source signals need not match the target waveforms for pre-training to help.
  • A testable extension would be to use the pre-trained encoder as a feature extractor for clustering or novelty detection on unseen radar emitters rather than only for linear classification; the t-SNE analysis hints that SSL improves class separation, but the paper does not quantify this with clustering metrics.
  • Because the best masking configurations were chosen per domain and shot count, an honest deployment protocol would fix the masking strategy and ratio on a separate validation set before touching the evaluation set; this is an editorial caution, not a claim in the paper.
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

2 major / 5 minor

Summary. The paper proposes a two-step self-supervised learning pipeline for few-shot radar signal recognition. In the first step, lightweight autoencoders (ResNet1D, MS-TCN, WaveNet) are pre-trained without labels on I/Q signals from radar, communication, or mixed RF domains using masked signal modelling with four masking strategies and varying masking ratios. In the second step, the pre-trained encoder is fine-tuned with a linear probe on 1-, 5-, or 10-shot radar data (RadChar-nShot) and evaluated on RadChar-Eval. The authors report that self-supervised pre-training improves 1-shot accuracy over from-scratch training by up to 17.5% relative (in-domain pre-training) and 16.31% relative (RadioML out-of-domain pre-training) for ResNet1D, and provide a benchmark of results across models, domains, and masking configurations.

Significance. The paper addresses a relevant problem and has several positive features: it releases a dataset family and benchmark, evaluates three lightweight architectures, and explores cross-domain pre-training, which is relatively uncommon in radar signal recognition. The empirical results, if confirmed under a sound selection protocol, would be practically useful. However, the central quantitative claim depends on how the 'optimal' masking strategy/ratio in Table 2 was chosen; the manuscript currently does not rule out test-set selection, which would make the reported gains upper bounds. The lack of repeated runs also makes it impossible to assess the stability of the improvements.

major comments (2)
  1. [Section 3.1, Section 3.2, Table 2] The central claim rests on comparing the 'optimal selection of Sm and Rm' for each model, domain, and shot against a no-SSL baseline. Section 3.1 explicitly states that 'no validation split is considered during fine-tuning' and that fine-tuning uses all available frames, while Section 3.2 reports the performance corresponding to 'the optimal selection of Sm and Rm used during pre-training.' The manuscript never states how this optimal selection was made. If the per-cell best entries in Table 2 and the curves in Fig. 2 were obtained by evaluating all four masking strategies and the displayed masking ratios against RadChar-Eval, then each reported accuracy is a maximum over a grid of configurations evaluated on the test set. Under that protocol, the headline 17.5% and 16.31% gains are upper bounds and would not support the claim that SSL pre-training is beneficial. The authors must state whether Sm/Rm were fixed a priori or selected on a held-out downstream validation set, and if the latter, the gains should be recomputed under that protocol.
  2. [Section 3.1, Table 2, Fig. 3] All experimental results are single runs with no error bars, confidence intervals, or significance tests. Given that the 1-shot fine-tuning set consists of only 205 frames, the observed differences between masking ratios and between models may be within run-to-run variance. At minimum, the headline claims for ResNet1D (17.5% and 16.31%) should be repeated over several random seeds and reported as mean plus/minus standard deviation, or accompanied by a suitable statistical test.
minor comments (5)
  1. [Section 2.3, Eq. (1)] The notation {X<Rm} is not defined; it should be stated explicitly that this denotes the indicator function of the event X<Rm, with X drawn uniformly from [0,1] as given in Eq. (2).
  2. [Abstract and title] The title contains an odd spacing in 'Adapta TION' that should be corrected, and the abstract should state whether the reported improvements are relative or absolute, since Table 2 implies relative improvements over the baseline.
  3. [Section 2.4, Table 1] The meaning of '1-shot corresponds to precisely 205 frames' is not explained; the paper should state how 205 follows from the number of classes and SNR levels in RadChar-nShot.
  4. [Section 3.2, Fig. 4] The t-SNE visualizations are qualitative; adding a quantitative cluster-quality metric or nearest-neighbor agreement would make the claim about improved feature separability more concrete.
  5. [Section 5, references] A few references have inconsistent formatting, for example [8] and [17], where the publisher or conference information is embedded in the author list; these should be normalized to the workshop style.

Circularity Check

1 steps flagged · score 6.0 of 10

Reported 'optimal' SSL gains are selected maxima on RadChar-Eval because no validation split exists for choosing Sm/Rm.

  1. fitted input called prediction [Section 3.1 (Training Details) and Section 3.2 (Table 2, Fig. 2)]
    "To maintain an even class distribution, no validation split is considered during fine-tuning. ... All models are evaluated on the RadChar-Eval dataset. ... The classification performance corresponding to the optimal selection of Sm and Rm used during pre-training is shown for each model."

    The paper selects the masking strategy Sm and masking ratio Rm as 'optimal' per model/domain/shot (Table 2) and then reports the resulting accuracy as the SSL result. Since no downstream validation split exists for fine-tuning, the only available selection criterion described is the RadChar-Eval test set (Section 3.1: 'All models are evaluated on the RadChar-Eval dataset'). Consequently each reported 'best' accuracy is a maximum over the evaluated Sm/Rm grid on the test distribution, and the headline 17.5% / 16.31% improvements are selected maxima rather than predictions from a fixed, pre-registered protocol.

full rationale

This is an empirical benchmark paper rather than a mathematical derivation, so most of the claimed improvements are measured test accuracies and are not circular in themselves. The main circularity-like step is the selection of masking hyperparameters Sm/Rm: Section 3.1 states that no validation split is considered during fine-tuning and that all models are evaluated on RadChar-Eval, while Section 3.2 reports the accuracy 'corresponding to the optimal selection' of Sm/Rm per model/domain/shot. Because no downstream validation set exists, the optimal selection is made using the test set, so each reported best accuracy is a maximum over the evaluated configuration grid and the headline 17.5%/16.31% gains are upper bounds selected on the test distribution. This fits the fitted-input-called-prediction pattern: the hyperparameters are fitted to the evaluation data and then reported as the SSL result. However, the paper's broader qualitative claims (SSL helps across several models/domains, t-SNE shows tighter clusters, out-of-domain RadioML also helps) retain independent empirical content, and the self-citations to the authors' RadChar dataset are used as data/benchmark rather than as a load-bearing theorem. The score is therefore 6 (partial circularity in the headline numbers) rather than 0 or 8.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities or parameters beyond hyperparameters; its contribution is empirical. Main free choices are masking strategy and ratio, whose tuning is not validation-based.

free parameters (2)
  • Masking ratio Rm = 0.1 to 0.9 depending on model, domain, and shot
    Chosen per experiment as the best over a grid of [0,1] values; selection appears to use test performance.
  • Masking strategy Sm = A, B, C, or D selected per model and domain
    Categorical design choice tuned per experiment; the per-domain optima in Table 2 vary and are not selected on a validation split.
assumptions (4)
  • domain assumption Mask signal reconstruction with L1 loss learns representations useful for classification
    This is the central SSL premise; the paper does not prove transfer, it demonstrates it empirically.
  • ad hoc to paper AWGN masking noise statistics from source domain (mu_train, sigma_train) are valid for corrupting signals across domains
    Equation (2) uses source-domain training statistics; no justification is given for why these are appropriate for target radar signals.
  • domain assumption RadChar-Eval is a representative, unbiased test set for radar signal recognition
    Evaluation relies on this dataset; no discussion of collection bias or class balance beyond counts is provided.
  • domain assumption Temporal resolution differences (tres) between source and target explain transfer effectiveness
    Offered as a post hoc hypothesis in Section 3.2, not tested directly.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation." pith.science (2026). https://pith.science/paper/243XX6PG

@misc{pith2026250103461,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/243XX6PG}},
  note         = {Machine review of arXiv:2501.03461}
}
read the original abstract

Radar signal recognition (RSR) plays a pivotal role in electronic warfare (EW), as accurately classifying radar signals is critical for informing decision-making. Recent advances in deep learning have shown significant potential in improving RSR in domains with ample annotated data. However, these methods fall short in EW scenarios where annotated radio frequency (RF) data are scarce or impractical to obtain. To address these challenges, we introduce a self-supervised learning (SSL) method which utilises masked signal modelling and RF domain adaption to perform few-shot RSR and enhance performance in environments with limited RF samples and annotations. We propose a two-step approach, first pre-training masked autoencoders (MAE) on baseband in-phase and quadrature (I/Q) signals from diverse RF domains, and then transferring the learned representations to the radar domain, where annotated data are scarce. Empirical results show that our lightweight self-supervised ResNet1D model with domain adaptation achieves up to a 17.5% improvement in 1-shot classification accuracy when pre-trained on in-domain signals (i.e., radar signals) and up to a 16.31% improvement when pre-trained on out-of-domain signals (i.e., comm signals), compared to its baseline without using SSL. We also present reference results for several MAE designs and pre-training strategies, establishing a new benchmark for few-shot radar signal classification.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    INTRODUCTION Radar signal recognition (RSR) is a crucial capability in cog- nitive electronic warfare (EW) [1], where accurate radar sig- nal classification is essential for informed decision-making in the battlefield. Recent progress in deep learning has demon- strated significant potential [2] in addressing RSR sub-tasks, such as automatic modulation cl...

    arXiv 2025

  2. [2]

    Two-Step Few-Shot Learning Our proposed SSL approach comprises two sequential steps

    PROPOSED METHOD 2.1. Two-Step Few-Shot Learning Our proposed SSL approach comprises two sequential steps. First, annotation-free pre-training of a masked autoencoder is conducted on a source RF domain (i.e., radar, comm, or a mixture of both). Then, the pre-trained encoder is fine-tuned on the target radar domain using a limited amount of anno- tated data...

    work page

  3. [3]

    Training Details We perform pre-training, fine-tuning, and model evaluation on a single Nvidia Tesla A100 GPU

    EXPERIMENTS 3.1. Training Details We perform pre-training, fine-tuning, and model evaluation on a single Nvidia Tesla A100 GPU. All models are trained with the Adam optimiser, where constant learning rates of 0.001 and 0.0001 are used for self-supervised pre-training and fine-tuning, respectively. For pre-training, we train each model for 100 epochs with ...

    work page

  4. [4]

    We also demonstrated the viability of RF domain adaptation for enhancing signal classification perfor- mance when no target domain data was used for pre-training

    CONCLUSION In this paper, we introduced MSM as an effective SSL method for few-shot RSR. We also demonstrated the viability of RF domain adaptation for enhancing signal classification perfor- mance when no target domain data was used for pre-training. Our results show that by optimally designing the masking method during pre-training, fine-tuned models ca...

    work page

  5. [5]

    Deep and Ensemble Learning to Win the Army RCO AI Signal Classification Challenge,

    Andres Vila, Donna Branchevsky, Kyle Logue, Sebastian Olsen, Esteban Valles, Darren Semmen, Alex Utter, and Eugene Grayver, “Deep and Ensemble Learning to Win the Army RCO AI Signal Classification Challenge,” in Proceedings of the 18th Python in Science Conference, 2019, pp. 21–26

    work page 2019

  6. [6]

    Karen Haigh and Julia Andrusenko, Cognitive Electronic Warfare: An Artificial Intelligence Approach, Artech House, 2021

    work page 2021

  7. [7]

    Deep Learning for Radar: A Survey,

    Zhe Geng, He Yan, Jindong Zhang, and Daiyin Zhu, “Deep Learning for Radar: A Survey,” IEEE Access, vol. 9, pp. 141800–141818, 2021

    work page 2021

  8. [8]

    Over-the-Air Deep Learning Based Radio Signal Classification,

    Timothy J. O’Shea, Tamoghna Roy, and T. Charles Clancy, “Over-the-Air Deep Learning Based Radio Signal Classification,” IEEE Journal of Selected Topics in Signal Processing, vol. 12, no. 1, pp. 168–179, 2018, IEEE

    work page 2018

Show all 41 references
  1. [9]

    Multi-stage Learning for Radar Pulse Activity Segmentation,

    Zi Huang, Akila Pemasiri, Simon Denman, Clinton Fookes, and Terrence Martin, “Multi-stage Learning for Radar Pulse Activity Segmentation,” in ICASSP 2024- 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 7340–7344

    2024

  2. [10]

    Multi-task Learning For Radar Signal Characterisation,

    Zi Huang, Akila Pemasiri, Simon Denman, Clinton Fookes, and Terrence Mar- tin, “Multi-task Learning For Radar Signal Characterisation,” in 2023 IEEE In- ternational Conference on Acoustics, Speech, and Signal Processing Workshops (ICASSPW), 2023, pp. 1–5

    2023

  3. [11]

    Automatic Modulation Classification: A Deep Architecture Survey,

    Thien Huynh-The, Quoc-Viet Pham, Toan-Van Nguyen, Thanh Thi Nguyen, Rukhsana Ruby, Ming Zeng, and Dong-Seong Kim, “Automatic Modulation Classification: A Deep Architecture Survey,” IEEE Access, vol. 9, pp. 142950– 142971, 2021

    2021

  4. [12]

    Deep Learning for Radar Signal Detection in the 3.5 GHz CBRS Band,

    Raied Caromi, Alex Lackpour, Kassem Kallas, Thao Nguyen, and Michael Souryal, “Deep Learning for Radar Signal Detection in the 3.5 GHz CBRS Band,” in 2021 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN). IEEE, 2021, pp. 1–8

    2021

  5. [13]

    Multi-task Learning Approach for Auto- matic Modulation and Wireless Signal Classification,

    Anu Jagannath and Jithin Jagannath, “Multi-task Learning Approach for Auto- matic Modulation and Wireless Signal Classification,” in ICC 2021-IEEE Inter- national Conference on Communications. 2021, pp. 1–7, IEEE

    2021

  6. [14]

    LSTM Framework for Classification of Radar and Communications Signals,

    Victoria Clerico, Jorge Gonz ´alez-L´opez, Gady Agam, and Jes ´us Grajal, “LSTM Framework for Classification of Radar and Communications Signals,” in 2023 IEEE Radar Conference (RadarConf23). IEEE, 2023, pp. 1–6

    2023

  7. [15]

    A Novel Approach to WaveNet Architecture for RF Signal Separation with Learnable Dila- tion and Data Augmentation,

    Yu Tian, Ahmed Alhammadi, Abdullah Quran, and Abubakar Sani Ali, “A Novel Approach to WaveNet Architecture for RF Signal Separation with Learnable Dila- tion and Data Augmentation,” in 2024 IEEE International Conference on Acous- tics, Speech, and Signal Processing Workshops (...

    2024

  8. [16]

    Automatic Radar Signal Detection and FFT Estimation using Deep Learning,

    Akila Pemasiri, Zi Huang, Fraser Williams, Ethan Goan, Simon Denman, Ter- rence Martin, and Clinton Fookes, “Automatic Radar Signal Detection and FFT Estimation using Deep Learning,” in 2024 17th International Conference on Sig- nal Processing and Communication System (ICSPCS)...

    2024

  9. [17]

    Learning Robust General Radio Signal Detection Using Computer Vision Methods,

    Tim O’Shea, Tamohgna Roy, and T Charles Clancy, “Learning Robust General Radio Signal Detection Using Computer Vision Methods,” in 2017 51st asilomar conference on signals, systems, and computers. IEEE, 2017, pp. 829–832

    2017

  10. [18]

    Accurate LPI Radar Waveform Recognition with CWD-TFA for Deep Convolutional Network,

    Thien Huynh-The, Van-Sang Doan, Cam-Hao Hua, Quoc-Viet Pham, Toan-Van Nguyen, and Dong-Seong Kim, “Accurate LPI Radar Waveform Recognition with CWD-TFA for Deep Convolutional Network,” IEEE Wireless Communications Letters, vol. 10, no. 8, pp. 1638–1642, 2021

    2021

  11. [19]

    Expert RF Feature Extrac- tion to Win the Army RCO AI Signal Classification Challenge,

    Kyle Logue, Esteban Valles, Andres Vila, Alex Utter, Darren Semmen, Eugene Grayver, Sebastian Olsen, and Donna Branchevsky, “Expert RF Feature Extrac- tion to Win the Army RCO AI Signal Classification Challenge,” in Proceedings of the 18th Python in Science Conference, 2019, pp. 8–14

    2019

  12. [20]

    BEiT: BERT Pre-training of Image Transformers,

    Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei, “BEiT: BERT Pre-training of Image Transformers,” arXiv preprint arXiv:2106.08254, 2021

    2021 arXiv

  13. [21]

    Deep Learning for RF Signal Classification in Unknown and Dynamic Spectrum Environments,

    Yi Shi, Kemal Davaslioglu, Yalin E Sagduyu, William C Headley, Michael Fowler, and Gilbert Green, “Deep Learning for RF Signal Classification in Unknown and Dynamic Spectrum Environments,” in 2019 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN). IEEE,...

    2019

  14. [22]

    Few-Shot Specific Emitter Identification Leveraging Neural Archi- tecture Search and Advanced Deep Transfer Learning,

    Weijie Zhang, Feng Shi, Qianyun Zhang, Yu Wang, Lantu Guo, Yun Lin, and Guan Gui, “Few-Shot Specific Emitter Identification Leveraging Neural Archi- tecture Search and Advanced Deep Transfer Learning,” IEEE Internet of Things Journal, 2024

    2024

  15. [23]

    Unsupervised Representation Learning of Structured Radio Communication Signals,

    Timothy J. O’Shea, Johnathan Corgan, and T. Charles Clancy, “Unsupervised Representation Learning of Structured Radio Communication Signals,” in 2016 First International Workshop on Sensing, Processing and Learning for Intelligent Machines (SPLINE). 2016, pp. 1–5, IEEE

    2016

  16. [24]

    Unsupervised Feature Learning and Automatic Modu- lation Classification Using Deep Learning Model,

    Afan Ali and Fan Yangyu, “Unsupervised Feature Learning and Automatic Modu- lation Classification Using Deep Learning Model,” Physical Communication, vol. 25, pp. 75–84, 2017

    2017

  17. [25]

    Frequency-Aware Masked Autoencoders for Multimodal Pretraining on Biosignals,

    Ran Liu, Ellen L Zippi, Hadi Pouransari, Chris Sandino, Jingping Nie, Hanlin Goh, Erdrin Azemi, and Ali Moin, “Frequency-Aware Masked Autoencoders for Multimodal Pretraining on Biosignals,” arXiv preprint arXiv:2309.05927, 2023

    2023 arXiv

  18. [26]

    Masked Autoencoders Are Scalable Vision Learners,

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Doll ´ar, and Ross Gir- shick, “Masked Autoencoders Are Scalable Vision Learners,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 16000–16009

    2022

  19. [27]

    Masked Autoencoders That Listen,

    Po-Yao Huang, Hu Xu, Juncheng Li, Alexei Baevski, Michael Auli, Wojciech Galuba, Florian Metze, and Christoph Feichtenhofer, “Masked Autoencoders That Listen,” Advances in Neural Information Processing Systems, vol. 35, pp. 28708– 28720, 2022

    2022

  20. [28]

    SSAST: Self- Supervised Audio Spectrogram Transformer,

    Yuan Gong, Cheng-I Lai, Yu-An Chung, and James Glass, “SSAST: Self- Supervised Audio Spectrogram Transformer,” in Proceedings of the AAAI Con- ference on Artificial Intelligence, 2022, vol. 36, pp. 10699–10709

    2022

  21. [29]

    MAEEG: Masked Auto-Encoder for EEG Representation Learning,

    Hsiang-Yun Sherry Chien, Hanlin Goh, Christopher M Sandino, and Joseph Y Cheng, “MAEEG: Masked Auto-Encoder for EEG Representation Learning,” arXiv preprint arXiv:2211.02625, 2022

    2022 arXiv

  22. [30]

    Deep Learning of Radio Frequency Fingerprints from Limited Samples by Masked Autoencoding,

    Keju Huang, Hui Liu, Pengjiang Hu, et al., “Deep Learning of Radio Frequency Fingerprints from Limited Samples by Masked Autoencoding,” IEEE Wireless Communications Letters, 2022

    2022

  23. [31]

    BERT: Pre- Training of Deep Bidirectional Transformers for Language Understanding,

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova, “BERT: Pre- Training of Deep Bidirectional Transformers for Language Understanding,” in Proceedings of naacL-HLT, 2019, vol. 1, p. 2

    2019

  24. [32]

    to construct our autoencoder as shown in Fig. 1. In the pre-training step, we utilise asymmetric masked autoencod- ing [21] whereby the model operates on a partially observed I/Q signal in the presence of additive white Gaussian noise (AWGN). We pre-train the model to reconstr...

  25. [33]

    A Transformer-Based Framework for Multivariate Time Series Representation Learning,

    George Zerveas, Srideepika Jayaraman, Dhaval Patel, Anuradha Bhamidipaty, and Carsten Eickhoff, “A Transformer-Based Framework for Multivariate Time Series Representation Learning,” in Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining, 2021, pp...

    2021

  26. [34]

    SimMIM: A Simple Framework for Masked Image Model- ing,

    Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu, “SimMIM: A Simple Framework for Masked Image Model- ing,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 9653–9663

    2022

  27. [35]

    Masked Autoencoders Enable Efficient Knowledge Distillers,

    Yutong Bai, Zeyu Wang, Junfei Xiao, Chen Wei, Huiyu Wang, Alan L Yuille, Yuyin Zhou, and Cihang Xie, “Masked Autoencoders Enable Efficient Knowledge Distillers,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 24256–24265

    2023

  28. [36]

    Few-Shot Specific Emitter Identification Using Asymmetric Masked Auto- Encoder,

    Zhisheng Yao, Xue Fu, Lantu Guo, Yu Wang, Yun Lin, Shengnan Shi, and Guan Gui, “Few-Shot Specific Emitter Identification Using Asymmetric Masked Auto- Encoder,” IEEE Communications Letters, 2023

    2023

  29. [37]

    Reducing the Dimensionality of Data with Neural Networks,

    Geoffrey E Hinton and Ruslan R Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” science, vol. 313, no. 5786, pp. 504–507, 2006

    2006

  30. [38]

    MS-TCN: Multi-Stage Temporal Convolu- tional Network for Action Segmentation,

    Yazan Abu Farha and Jurgen Gall, “MS-TCN: Multi-Stage Temporal Convolu- tional Network for Action Segmentation,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition, 2019, pp. 3575–3584

    2019

  31. [39]

    WaveNet: A Generative Model for Raw Audio,

    Aaron Van Den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, Koray Kavukcuoglu, et al., “WaveNet: A Generative Model for Raw Audio,” arXiv preprint arXiv:1609.03499, vol. 12, 2016

    2016 arXiv

  32. [40]

    The Data-Driven Radio Frequency Signal Separation Challenge,

    Tejas Jayashankar, Binoy Kurien, Alejandro Lancho, Gary CF Lee, Yury Polyan- skiy, Amir Weiss, and Gregory W Wornell, “The Data-Driven Radio Frequency Signal Separation Challenge,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process.(ICASSP), 2024

    2024

  33. [41]

    Generative Pretraining from Pixels,

    Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever, “Generative Pretraining from Pixels,” inInternational confer- ence on machine learning. PMLR, 2020, pp. 1691–1703

    2020

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.