Pith. sign in

REVIEW 4 major objections 5 minor 64 references

The paper claims that a self-supervised masked autoencoder reconstructing both amplitude and phase of WiFi CSI, guided by an adaptive masking policy, can beat fully supervised classifiers on one benchmark and surpass all compared self-super

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 ↗

A dual-stream masked autoencoder with adaptive masking and Barlow Twins alignment learns WiFi CSI representations that beat prior self-supervised baselines and, on SignFi, a fully supervised model.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection A sensible CSI-HAR SSL recipe with an unproven headline: the pretraining/test split ambiguity and missing error bars make the 'beats supervised' claim unverifiable until the authors clarify. the 4 major comments →

arxiv 2512.04723 v2 pith:EFKSRU4W submitted 2025-12-04 eess.SP

CIG-MAE: Cross-Modal Information-Guided Masked Autoencoder for Self-Supervised WiFi Sensing

Gang Liu , Yanling Hao , Yixuan Zou This is my paper
classification eess.SP
keywords WiFi sensingChannel State Information (CSI)self-supervised learningmasked autoencoderhuman activity recognitioncross-modal representation learningadaptive maskingBarlow Twins regularizer
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

This paper sets out to show that self-supervised pretraining on unlabeled WiFi signals can replace most labeled data collection for human action recognition. The proposed method, CIG-MAE, is a masked autoencoder that reconstructs both the amplitude and phase of WiFi Channel State Information (CSI) through two symmetric convolutional streams, using a policy network to concentrate its tiny visible budget on time–frequency patches that are hardest to reconstruct, and a Barlow Twins regularizer to align the two modalities. Under standard k-shot linear probing, frozen features from this pretraining beat all compared self-supervised baselines on three public benchmarks, and on the sign-language dataset they also exceed a fully supervised classifier trained on all labels. If these results hold, WiFi sensing can be bootstrapped cheaply from unlabeled signal data, and the small CNN backbone makes deployment on resource-constrained edge devices realistic.

Core claim

The paper's central claim is that the physical structure of CSI—paired amplitude and phase, non-uniform information density on the time–frequency plane, and heavy noise—should dictate the self-supervised objective. CIG-MAE reconstructs both modalities from 5% visible patches, using an adaptive masking policy that keeps high-error patches visible and a Barlow Twins term that aligns the two encoders' features. On three public datasets the frozen representations reach 98.55%, 63.50%, and 40.10% accuracy under 1- or 10-shot linear probing, surpassing all SSL baselines and, on the sign-language set, the fully supervised model. This is achieved with a 53.46 MB CNN backbone at 0.14 GFLOPs, about 10

What carries the argument

The load-bearing object is the Adaptive Information-Guided Masking (AIM) policy in combination with a symmetric dual-stream autoencoder. AIM is a lightweight Transformer policy network that assigns each time–frequency patch a visibility probability, then uses Gumbel-TopK sampling to select the small visible set (5%). It is trained by a policy-gradient loss whose reward is the mean absolute reconstruction error of the masked patches, with a stop-gradient on that reward so the reconstruction backbone cannot inflate errors to please the policy. The dual-stream CNN encoder-decoder reconstructs amplitude and phase separately, and a Barlow Twins projector aligns their latent spaces by pushing the

Load-bearing premise

The load-bearing premise is that a patch's mean reconstruction error under the current model is a faithful proxy for task-relevant information density in CSI — if high error is actually caused by noise spikes, phase wrapping, or unreconstructable artifacts, AIM will preserve noise rather than activity and the reported gains will not transfer.

What would settle it

Pretrain CIG-MAE on a CSI dataset with known ground-truth activity regions (or on synthetic signals with a planted activity mask and noise), then compare the patches AIM marks visible against the true activity mask. If the visible set does not preferentially overlap the activity regions while accuracy still beats random masking, the information-density reward is not the causal mechanism; alternatively, replace the reconstruction-error reward with a noise-only proxy and show the accuracy gains persist.

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

If this is right

  • Unlabeled WiFi data can replace most labeled collection: frozen CIG-MAE features beat the fully supervised baseline on the sign-language dataset with a single labeled example per class.
  • The same pretrained representations transfer across tasks (sign language, daily activities, human interactions) and outperform prior SSL methods on all three, supporting the idea of a general CSI pretraining recipe.
  • Small CNN backbones are sufficient: 53.46 MB parameters, 0.14 GFLOPs, and 131.97 MB memory, which makes real-time inference on commercial access points plausible (the ViT equivalent costs roughly 608 MB and 129.98 GFLOPs).
  • Pretraining benefits from more unlabeled data, especially in noisy scenarios: on the two more difficult benchmarks, raising pretraining data from 50% to 80% improves F1 substantially, and the gains widen with more labeled shots.
  • Adaptive masking is a major source of gain: replacing AIM with random masking drops accuracy by up to 5.42 points on the human-interaction dataset, so concentrating the visibility budget on high-error patches is what drives the improvement.

Where Pith is reading between the lines

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

  • The AIM policy's learned visibility probabilities are a ready-made saliency map of where activity lives in the time–frequency plane; the paper does not exploit them beyond masking, but they could be reused as a free interpretability or weak-localization signal.
  • The reconstruction-error-as-information-density hypothesis is modality-agnostic, so the same dual-stream adaptive-masking recipe could transfer to radar micro-Doppler, mmWave, or sonar sensing, which share noisy time–frequency structure and suffer from physically meaningless augmentations.
  • A direct causal test would compare AIM's chosen patches with ground-truth activity masks on synthetic or labeled CSI; if they do not preferentially overlap while accuracy still beats random masking, the information-density reward is not the mechanism.
  • The stop-gradient policy-gradient design resembles a two-player game; a simpler reward based on reconstruction-loss gradient magnitude might achieve similar gains without a separate policy network, which is an easily testable simplification.
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 / 5 minor

Summary. The paper proposes CIG-MAE, a self-supervised learning method for WiFi CSI human activity recognition. It uses a symmetric dual-stream CNN autoencoder to reconstruct masked amplitude and phase patches, an Adaptive Information-Guided Masking (AIM) policy that selects visible patches based on reconstruction error, and a Barlow Twins regularizer to align the two modality streams. The authors evaluate under k-shot linear probing on three public datasets (SignFi, NLOS, HTHI), reporting that CIG-MAE outperforms AutoFi, AutoSen, and CAPC and even exceeds a fully supervised baseline on SignFi. The paper also compares CNN and ViT backbones, claiming large accuracy and efficiency advantages for the CNN.

Significance. If the reported results are reproducible and the pretraining protocol is clean, the paper would provide a practical SSL recipe for small CSI datasets that needs no contrastive augmentations, no large batches, and runs on a lightweight CNN, which is appealing for edge deployment. The combination of masked reconstruction, adaptive masking, and cross-modal regularization is reasonable. However, the central empirical claims rest on three load-bearing points that currently need attention: the pretraining data split must be shown to exclude test samples; the AIM loss in Eq. (4) needs a correct derivation or an honest heuristic label; and the reported point estimates need variance information before 'consistently outperforms' can be assessed.

major comments (4)
  1. [Section IV-A and Fig. 7] The manuscript states 'Datasets are stratified (8:2 train/test split)' but never says whether the 300-epoch pretraining phase uses only the training split or the full dataset. Figure 7 varies 'unlabeled pre-training data ratio (50%, 65%, 80%)' in a way that appears to sample from the whole dataset, with no exclusion of test samples. If the test split is included in pretraining, the frozen encoder has seen the exact test CSI tensors, making the reported 98.55% vs 96.38% SignFi comparison and the NLOS/HTHI gains transductive rather than evidence of generalization. This is the single most important issue: please state explicitly which split was used for pretraining, and if the full dataset was used, rerun all methods with train-only pretraining and report corrected numbers.
  2. [Eq. (4), Sec. III-B] The AIM objective is described as a 'policy gradient loss', but Eq. (4) is not the gradient of an expected reward under the Gumbel-TopK sampling distribution. p_i is defined as the probability that patch i is visible, and the sum runs over the masked set M. For a Gumbel-TopK draw, the log-probability of the sampled partition involves the selected visible set and the sequential normalization terms; log p_i for i in M is not the score function of the sampled mask. Thus minimizing Eq. (4) is not guaranteed to maximize E[R(M)] as claimed. Please either provide a correct policy-gradient derivation or explicitly state that Eq. (4) is a heuristic reward-weighted likelihood used to bias the mask distribution.
  3. [Tables II-V and Figs. 6-7] All experimental results are single point estimates with no error bars, confidence intervals, or repeated-seed statistics. The claims include 'consistently outperforms' and 'surpasses a fully supervised baseline', with margins as small as 2.17% on SignFi (98.55 vs 96.38). Given stochastic pretraining, Gumbel-TopK sampling, and k-shot linear probing, the observed margins could easily lie within run-to-run variance. Please report mean and standard deviation over at least 5 random seeds (or equivalently robust bootstrap intervals) for the main tables and key sensitivity figures.
  4. [Sec. III-B, AIM assumption] The load-bearing premise is that 'patches that are difficult to reconstruct from their context contain higher information density'. The entire AIM mechanism and the reported improvements depend on this assumption. The paper does not provide evidence that high reconstruction error correlates with task-relevant activity rather than with hardware noise spikes, phase wrapping, or unreconstructable artifacts. Beyond the qualitative visualizations in Fig. 4, please add a quantitative test: e.g., compare AIM-selected patches against activity-labeled time-frequency regions, or run an ablation where the reward is replaced by a physical saliency measure (such as local SNR or Doppler energy).
minor comments (5)
  1. [Sec. III-B and Ref. [49]] The paper says 'we incorporate the AIM strategy [49]' but later lists AIM as a new contribution. Please clarify explicitly what is borrowed from AdaMAE and what is new, especially the use of reconstruction error as reward and the stop-gradient decoupling.
  2. [Sec. III-B, Eq. (3)] The softmax over all L patches defines a categorical distribution over patches, but Gumbel-TopK sampling draws a fixed-size subset. The relationship between p_i and the probability of a particular visible set should be stated precisely.
  3. [Table I] The Input row says 'CSI:3(6)×30×200'. Please explain the 3(6) notation (number of antennas per dataset) in the table caption or text.
  4. [Sec. IV-E, Table IV] The comparison 'MAE (w/o Norm.)' vs normalized variants is interesting, but the text says normalization compresses dynamic range. Since z-score normalization is already applied to both modalities before the loss (Sec. IV-A), please clarify how 'w/o Norm.' differs from the preprocessing normalization.
  5. [General] Figure 1 and Figure 4 use color heatmaps without colorbar units. Adding a shared scale would help readers interpret reconstruction error and mask density.

Circularity Check

0 steps flagged

No load-bearing circularity; the central linear-probe evaluation is independent of the training objective, and the single self-citation is a non-load-bearing baseline.

full rationale

Central claim — CIG-MAE beats SOTA SSL methods and the supervised baseline under k-shot linear probing — is not derived from its own inputs by construction. The evaluation is an independent linear probe on frozen encoder features with held-out labels, while the pretraining objective (Eqs. 4, 5, 8, 9) is label-free. The AIM reward (Eq. 4) is the mask-conditional reconstruction error of the very backbone being trained; this is an online RL co-adaptation loop, not a logical reduction, and the stopgrad operator plus the decoupled updates (Algorithm 1, lines 16–21) explicitly prevent the backbone from inflating the reward. 'Information density' is stated as a hypothesis ('patches that are difficult to reconstruct from their context contain higher information density', Sec. III-B) and is tested on held-out labels by the w/o AIM ablation (Table III), so it is a falsifiable assumption, not a circular definition. The only overlapping-author citation, AutoSen [14] (co-author Y. Hao), is used as a baseline and outperformed (Table II) and critiqued, so it is not load-bearing. Flagged as correctness risk rather than circularity: Sec. IV-A states 'Datasets are stratified (8:2 train/test split)' but never states whether the 300-epoch pretraining uses only the training split; Fig. 7's sweep of unlabeled pretraining ratios (50%, 65%, 80%) could be read as drawing from the full dataset. If test CSI tensors were included in pretraining, the reported 98.55%/63.50%/40.10% would be partially transductive, not predictive; the authors must specify the pretraining split and release data indices. The per-dataset BT projection width (1024 vs 5096, Sec. IV-A) was evidently selected against these same test sets, which weakens but does not circularize the comparison.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 0 invented entities

All core hyperparameters are chosen on the same three datasets (often with knowledge of the evaluation protocol), so the empirical gains are conditional on these choices. The physically motivated assumptions — that reconstruction difficulty tracks action-relevant information, that amplitude and phase are complementary, and that CNN locality fits CSI — are plausible but not independently verified. The AIM policy is trained with a self-generated reward (reconstruction error), so its success is an empirical claim, not a guaranteed property.

free parameters (6)
  • mask ratio rho = 0.95
    Chosen from sensitivity analysis (Fig. 6d); central to the reconstruction task.
  • BT loss weight w_bt = 0.2
    Selected via sensitivity analysis (Fig. 6c).
  • AIM loss weight w_aim = 1e-4
    Fixed in Sec. IV-A; no sensitivity shown.
  • BT projection head width = 1024 (SignFi/HTHI), 5096 (NLOS)
    Per-dataset choice; the paper says NLOS requires larger capacity (Sec. IV-E).
  • AIM feature dimension x_m = 256
    Selected via sensitivity analysis (Fig. 6a).
  • patch size (s_p, t_p) = (3,5)
    Set in Sec. IV-A to align with first conv kernel/stride; not ablated.
axioms (6)
  • ad hoc to paper Reconstruction error equals information density
    Sec. III-B hypothesis: 'patches that are difficult to reconstruct ... contain higher information density'; no independent ground-truth validation.
  • domain assumption Amplitude and phase are complementary and both needed for HAR
    Sec. I and III-A; physically motivated, but the paper does not quantify per-modality discriminative contribution except via dual-stream ablation.
  • domain assumption Barlow Twins decorrelation preserves task-relevant information
    Borrowed from [24]; the paper assumes redundancy reduction does not discard action-discriminative features.
  • domain assumption CNN inductive bias is suitable for CSI time-frequency patterns
    Sec. III-C; supported by the ViT comparison, but only on these three datasets.
  • domain assumption Policy gradient with Gumbel-TopK and stop-gradient converges to a useful masking policy
    Sec. III-B/Alg. 1; no convergence guarantee or reward stability analysis.
  • domain assumption Z-score normalization and phase calibration preserve label-relevant structure
    Sec. IV-A; preprocessing changes the raw signal and is assumed not to remove discriminative information.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of CIG-MAE: Cross-Modal Information-Guided Masked Autoencoder for Self-Supervised WiFi Sensing." pith.science (2026). https://pith.science/paper/EFKSRU4W

@misc{pith2026251204723,
  author       = {Pith},
  title        = {Pith review of: CIG-MAE: Cross-Modal Information-Guided Masked Autoencoder for Self-Supervised WiFi Sensing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EFKSRU4W}},
  note         = {Machine review of arXiv:2512.04723}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Human Action Recognition using WiFi Channel State Information (CSI) has emerged as an attractive alternative to vision-based methods due to its ubiquity, device-agnostic nature, and inherent privacy-preserving capabilities. However, the high cost of manual annotation and the limited scale of publicly available CSI datasets restrict the performance of supervised approaches. Self-supervised learning (SSL) offers a promising avenue, but existing contrastive paradigms rely on data augmentations that conflict with the physical semantics of radio signals and require large-batch training, making them poorly suited for CSI. To overcome these challenges, we introduce CIG-MAE -- a Cross-modal Information-Guided Masked Autoencoder -- that reconstructs both the amplitude and phase of CSI using a symmetric dual-stream architecture with a high masking ratio. Specifically, we propose an Adaptive Information-Guided Masking strategy that dynamically allocates attention to time-frequency regions with high information density to improve learning efficiency, and incorporate a Barlow Twins regularizer to align cross-modal representations without negative samples. Experiments on three public datasets show that CIG-MAE consistently outperforms SOTA SSL methods and even surpasses a fully supervised baseline, demonstrating superior data efficiency, robustness, and representation generalization.

Figures

Figures reproduced from arXiv: 2512.04723 by Gang Liu, Yanling Hao, Yixuan Zou.

Figure 1
Figure 1. Figure 1: Reconstruction error heatmaps for amplitude and phase on the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Illustration of five visualization/augmentation operators applied to [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Schematic overview of the CIG-MAE framework. The architecture integrates three core modules corresponding to Section III: (1) [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Demonstration of AIM’s visibility allocation. Compared to ran [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Demonstration of the alignment and decorrelation effect of BT. After [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: First, performance monotonically improves with labeled samples (k) across all settings. This confirms that CIG-MAE learns a meaningful, linearly separable feature space where the classifier can effectively leverage additional supervision to refine decision boundaries. Second, the impact of pre-training data volume depends heavily on dataset complexity. On the clean SignFi dataset, the gain from 50% to 80% … view at source ↗
Figure 6
Figure 6. Figure 6: Sensitivity analysis of CIG-MAE’s key hyperparameters across [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Impact of unlabeled pre-training data size and labeled fine-tuning [PITH_FULL_IMAGE:figures/full_fig_p009_7.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

64 extracted references · 24 linked inside Pith

  1. [1]

    Single-Target Real-Time Passive WiFi Tracking,

    Z. Wang, J. A. Zhang, M. Xu, and Y . J. Guo, “Single-Target Real-Time Passive WiFi Tracking,”IEEE Trans. Mob. Comput., vol. 22, no. 6, pp. 3724–3742, Jun. 2023

    2023

  2. [2]

    WiFi-Based Real-Time Breath- ing and Heart Rate Monitoring During Sleep,

    Y . Gu, X. Zhang, Z. Liu, and F. Ren, “WiFi-Based Real-Time Breath- ing and Heart Rate Monitoring During Sleep,” inProc. IEEE Global Commun. Conf. (GLOBECOM), Dec. 2019, pp. 1–6

    2019

  3. [3]

    BodyCompass: Monitoring sleep posture with wireless signals,

    S. Yue, Y . Yang, H. Wang, H. Rahul, and D. Katabi, “BodyCompass: Monitoring sleep posture with wireless signals,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 4, no. 2, p. 66, Jun. 2020

    2020

  4. [4]

    RT-Fall: A real-time and contactless fall detection system with commodity WiFi devices,

    H. Wang, D. Zhang, Y . Wang, J. Ma, Y . Wang, and S. Li, “RT-Fall: A real-time and contactless fall detection system with commodity WiFi devices,”IEEE Trans. Mob. Comput., vol. 16, no. 2, pp. 511–526, Feb. 2017

    2017

  5. [5]

    Emergency Semantic Feature Vector Extraction From WiFi Signals for In-Home Monitoring of Elderly,

    L. Guo, Z. Lu, S. Zhou, X. Wen, and Z. He, “Emergency Semantic Feature Vector Extraction From WiFi Signals for In-Home Monitoring of Elderly,”IEEE J. Sel. Topics Signal Process., vol. 15, no. 6, pp. 1423–1438, Nov. 2021

    2021

  6. [6]

    Wi-fi can do more: Toward ubiquitous wireless sensing,

    C. Wu, B. Wang, O. C. Au, and K. J. R. Liu, “Wi-fi can do more: Toward ubiquitous wireless sensing,”IEEE Communications Standards Magazine, vol. 6, no. 2, pp. 42–49, Jun. 2022

    2022

  7. [7]

    WiFi Sensing with Channel State Information: A Survey,

    Y . Ma, G. Zhou, and S. Wang, “WiFi Sensing with Channel State Information: A Survey,”ACM Comput. Surv., vol. 52, no. 3, p. 46, Jun. 2019

    2019

  8. [8]

    Transfer Learning in Human Activity Recognition: A Survey,

    S. G. Dhekane and T. Ploetz, “Transfer Learning in Human Activity Recognition: A Survey,” 2024, arXiv preprint arXiv:2401.10185

    Pith/arXiv arXiv 2024

  9. [9]

    Towards Environment Independent Device Free Human Activity Recognition,

    W. Jiang, C. Miao, F. Ma, S. Yao, Y . Wang, Y . Yuan, H. Xue, C. Song, X. Ma, D. Koutsonikolas, W. Xu, and L. Su, “Towards Environment Independent Device Free Human Activity Recognition,” inProc. 24th Annu. Int. Conf. Mobile Comput. Netw., 2018, pp. 289–304

    2018

  10. [10]

    A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends,

    J. Gui, T. Chen, J. Zhang, Q. Cao, Z. Sun, H. Luo, and D. Tao, “A Survey on Self-supervised Learning: Algorithms, Applications, and Future Trends,” 2024, arXiv preprint arXiv:2301.05712

    Pith/arXiv arXiv 2024

  11. [11]

    Self-Supervised Learning for WiFi CSI-Based Human Activity Recognition: A Systematic Study,

    K. Xu, J. Wang, H. Zhu, and D. Zheng, “Self-Supervised Learning for WiFi CSI-Based Human Activity Recognition: A Systematic Study,” 2023, arXiv preprint arXiv:2308.02412

    Pith/arXiv arXiv 2023

  12. [12]

    Diffusion- Model-Based Contrastive Learning for Human Activity Recognition,

    C. Xiao, Y . Han, W. Yang, Y . Hou, F. Shi, and K. Chetty, “Diffusion- Model-Based Contrastive Learning for Human Activity Recognition,” IEEE Internet Things J., vol. 11, no. 20, pp. 33 525–33 536, Oct. 2024

    2024

  13. [13]

    Device- Free Wireless Sensing With Few Labels Through Mutual Information Maximization,

    B. Chen, J. Wang, Y . Lv, Q. Gao, M. Pan, and Y . Fang, “Device- Free Wireless Sensing With Few Labels Through Mutual Information Maximization,”IEEE Internet Things J., vol. 11, no. 6, pp. 10 513– 10 524, Mar. 2024

    2024

  14. [14]

    Autosen: improving automatic WiFi human sensing through cross-modal autoencoder,

    Q. Gao, Y . Hao, and Y . Liu, “Autosen: improving automatic WiFi human sensing through cross-modal autoencoder,” 2024, arXiv preprint arXiv:2401.05440

    Pith/arXiv arXiv 2024

  15. [15]

    BERT: Pre- training of deep bidirectional transformers for language understand- ing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre- training of deep bidirectional transformers for language understand- ing,” 2019, arXiv preprint arXiv:1810.04805

    Pith/arXiv arXiv 2019

  16. [16]

    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,” 2021, arXiv preprint arXiv:2111.06377

    Pith/arXiv arXiv 2021

  17. [17]

    BEiT: BERT pre-training of image transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “BEiT: BERT pre-training of image transformers,” 2022, arXiv preprint arXiv:2106.08254

    Pith/arXiv arXiv 2022

  18. [18]

    Self-Supervised Speech Representation Learning: A Review,

    A. Mohamed, H.-y. Lee, L. Borgholt, J. D. Havtorn, J. Edin, C. Igel, K. Kirchhoff, S.-W. Li, K. Livescu, L. Maaløe, T. N. Sainath, and S. Watanabe, “Self-Supervised Speech Representation Learning: A Review,”IEEE J. Sel. Topics Signal Process., vol. 16, no. 6, pp. 1179– 1210, Oct. 2022

    2022

  19. [19]

    MaskCAE: Masked convolutional autoencoder via sensor data re- construction for self-supervised human activity recognition,

    D. Cheng, L. Zhang, L. Qin, S. Wang, H. Wu, and A. Song, “MaskCAE: Masked convolutional autoencoder via sensor data re- construction for self-supervised human activity recognition,”IEEE J. Biomed. Health Inform., vol. 28, no. 5, pp. 2687–2698, May 2024

    2024

  20. [20]

    Sense and Learn: Self- supervision for omnipresent sensors,

    A. Saeed, V . Ungureanu, and B. Gfeller, “Sense and Learn: Self- supervision for omnipresent sensors,”Mach. Learn. Appl., vol. 6, p. 100152, 2021

    2021

  21. [21]

    RF-URL: unsupervised representation learning for RF sensing,

    R. Song, D. Zhang, Z. Wu, C. Yu, C. Xie, S. Yang, Y . Hu, and Y . Chen, “RF-URL: unsupervised representation learning for RF sensing,” in Proc. 28th Annu. Int. Conf. Mobile Comput. Netw., 2022, pp. 282– 295

    2022

  22. [22]

    Federated Self- Supervised Learning of Multisensor Representations for Embedded Intelligence,

    A. Saeed, F. D. Salim, T. Ozcelebi, and J. Lukkien, “Federated Self- Supervised Learning of Multisensor Representations for Embedded Intelligence,”IEEE Internet Things J., vol. 8, no. 2, pp. 1030–1040, Jan. 2021

    2021

  23. [23]

    Momentum Con- trast for Unsupervised Visual Representation Learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum Con- trast for Unsupervised Visual Representation Learning,” 2020, arXiv preprint arXiv:1911.05722

    Pith/arXiv arXiv 2020

  24. [24]

    Barlow twins: Self-supervised learning via redundancy reduction,

    J. Zbontar, L. Jing, I. Misra, Y . LeCun, and S. Deny, “Barlow twins: Self-supervised learning via redundancy reduction,”arXiv preprint arXiv:2103.03230, 2021

    Pith/arXiv arXiv 2021

  25. [25]

    Vicreg: Variance-invariance- covariance regularization for self-supervised learning,

    A. Bardes, J. Ponce, and Y . LeCun, “Vicreg: Variance-invariance- covariance regularization for self-supervised learning,”arXiv preprint arXiv:2105.04906, 2022

    Pith/arXiv arXiv 2022

  26. [26]

    A Simple Framework for Contrastive Learning of Visual Representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A Simple Framework for Contrastive Learning of Visual Representations,” 2020, arXiv preprint arXiv:2002.05709

    Pith/arXiv arXiv 2020

  27. [27]

    Self-Supervised WiFi-Based Activity Recognition,

    M. J. Bocus, H.-S. Lau, R. McConville, R. J. Piechocki, and R. Santos- Rodriguez, “Self-Supervised WiFi-Based Activity Recognition,” in Proc. IEEE Globecom Workshops (GC Wkshps), Dec. 2022, pp. 552– 557

    2022

  28. [28]

    Con- trastive Self-Supervised Representation Learning for Sensing Signals from the Time-Frequency Perspective,

    D. Liu, T. Wang, S. Liu, R. Wang, S. Yao, and T. Abdelzaher, “Con- trastive Self-Supervised Representation Learning for Sensing Signals from the Time-Frequency Perspective,” inProc. Int. Conf. Comput. Commun. Netw. (ICCCN), Jul. 2021, pp. 1–10

    2021

  29. [29]

    Dual-Stream Contrastive Learning for Channel State Information Based Human Activity Recognition,

    K. Xu, J. Wang, L. Zhang, H. Zhu, and D. Zheng, “Dual-Stream Contrastive Learning for Channel State Information Based Human Activity Recognition,”IEEE J. Biomed. Health Inform., vol. 27, no. 1, pp. 329–338, Jan. 2023

    2023

  30. [30]

    WiFiAct: Enhancing Human Sensing Through Environment Robust Preprocessing and Bayesian Self-Supervised Learning,

    N. Lyons, A. Santra, V . K. Ramanna, K. Uln, R. Taori, and A. Pandey, “WiFiAct: Enhancing Human Sensing Through Environment Robust Preprocessing and Bayesian Self-Supervised Learning,” in2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, April 2024, pp. 13 391–13 395

    2024

  31. [31]

    AutoFi: Toward au- tomatic Wi-Fi human sensing via geometric self-supervised learning,

    J. Yang, X. Chen, H. Zou, D. Wang, and L. Xie, “AutoFi: Toward au- tomatic Wi-Fi human sensing via geometric self-supervised learning,” IEEE Internet Things J., vol. 10, no. 8, pp. 7416–7425, Apr. 2023

    2023

  32. [32]

    Representation Learn- ing with Contrastive Predictive Coding,

    A. van den Oord, Y . Li, and O. Vinyals, “Representation Learn- ing with Contrastive Predictive Coding,” 2019, arXiv preprint arXiv:1807.03748

    Pith/arXiv arXiv 2019

  33. [33]

    Self-supervised learning for time series analysis: Taxonomy, progress, and prospects,

    K. Zhang, Q. Wen, C. Zhang, R. Cai, M. Jin, Y . Liu, J. Y . Zhang, Y . Liang, G. Pang, D. Song, and S. Pan, “Self-supervised learning for time series analysis: Taxonomy, progress, and prospects,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 10, pp. 6775–6794, 2024

    2024

  34. [34]

    Contrastive Predic- tive Coding for Human Activity Recognition,

    H. Haresamudram, I. Essa, and T. Ploetz, “Contrastive Predic- tive Coding for Human Activity Recognition,” 2020, arXiv preprint arXiv:2012.05333

    Pith/arXiv arXiv 2020

  35. [35]

    Context-Aware Predictive Coding: A Representation Learning Framework for WiFi Sensing,

    B. Barahimi, H. Tabassum, M. Omer, and O. Waqar, “Context-Aware Predictive Coding: A Representation Learning Framework for WiFi Sensing,” 2024, arXiv preprint arXiv:2410.01825

    Pith/arXiv arXiv 2024

  36. [36]

    Strohmayer and M

    J. Strohmayer and M. Kampel,Data Augmentation Techniques for Cross-Domain WiFi CSI-Based Human Activity Recognition, 2024, pp. 42–56

    2024

  37. [37]

    Simple and Effective Augmentation Methods for CSI Based Indoor Localization,

    O. G. Serbetci, J.-H. Lee, D. Burghal, and A. F. Molisch, “Simple and Effective Augmentation Methods for CSI Based Indoor Localization,” 2023, arXiv preprint arXiv:2211.10790

    Pith/arXiv arXiv 2023

  38. [38]

    RFBoost: Understanding and boosting deep WiFi sensing via physical data augmentation,

    W. Hou and C. Wu, “RFBoost: Understanding and boosting deep WiFi sensing via physical data augmentation,”Proc. ACM Interact., Mobile, Wearable Ubiquitous Technol., vol. 8, no. 2, pp. 1–26, May 2024

    2024

  39. [39]

    ImageNet: A large-scale hierarchical image database,

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A large-scale hierarchical image database,” inProc. IEEE Conf. Comput. Vis. Pattern Recognit., Jun. 2009, pp. 248–255

    2009

  40. [40]

    A Tutorial-cum-Survey on Self-Supervised Learning for Wi-Fi Sensing: Trends, Challenges, and Outlook,

    A. Y . Radwan, M. Yildirim, N. Hasanzadeh, H. Tabassum, and S. Valaee, “A Tutorial-cum-Survey on Self-Supervised Learning for Wi-Fi Sensing: Trends, Challenges, and Outlook,”IEEE Commun. Surv. Tuts., pp. 1–1, 2025

    2025

  41. [41]

    Understanding Masked Image Model- ing via Learning Occlusion Invariant Feature,

    X. Kong and X. Zhang, “Understanding Masked Image Model- ing via Learning Occlusion Invariant Feature,” 2022, arXiv preprint arXiv:2208.04164

    Pith/arXiv arXiv 2022

  42. [42]

    Spatial-Temporal Masked Autoencoder for Multi-Device Wearable Human Activity Recognition,

    S. Miao, L. Chen, and R. Hu, “Spatial-Temporal Masked Autoencoder for Multi-Device Wearable Human Activity Recognition,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 7, no. 4, p. 172, Jan. 2024

    2024

  43. [43]

    MaskFi: Unsupervised learning of WiFi and vision representations for multimodal human activity recognition,

    J. Yang, S. Tang, Y . Xu, Y . Zhou, and L. Xie, “MaskFi: Unsupervised learning of WiFi and vision representations for multimodal human activity recognition,” 2024, arXiv preprint arXiv:2402.19258

    Pith/arXiv arXiv 2024

  44. [44]

    SiFall: Practical online fall detection with RF sensing,

    S. Ji, Y . Xie, and M. Li, “SiFall: Practical online fall detection with RF sensing,” inProc. 20th ACM Conf. Embedded Netw. Sensor Syst., Nov. 2022, pp. 563–577

    2022

  45. [45]

    Efficient Training of Visual Transformers with Small Datasets,

    Y . Liu, E. Sangineto, W. Bi, N. Sebe, B. Lepri, and M. D. Nadai, “Efficient Training of Visual Transformers with Small Datasets,” 2021, arXiv preprint arXiv:2106.03746

    Pith/arXiv arXiv 2021

  46. [46]

    Kakogeorgiou, S

    I. Kakogeorgiou, S. Gidaris, B. Psomas, Y . Avrithis, A. Bursuc, K. Karantzalos, and N. Komodakis,What to Hide from Your Students: Attention-Guided Masked Image Modeling, 2022, pp. 300–318

    2022

  47. [47]

    SLNet: A spectrogram learning neural network for deep wireless sensing,

    Z. Yang, Y . Zhang, K. Qian, and C. Wu, “SLNet: A spectrogram learning neural network for deep wireless sensing,” inProc. 20th USENIX Symp. Netw. Syst. Des. Implementation (NSDI 23), Apr. 2023, pp. 1221–1236

    2023

  48. [48]

    Subcar- rier selection for efficient CSI-based indoor localization,

    Y . Taso, S.-C. Yeh, Y .-Y . Liang, C.-H. Wang, and S.-H. Fang, “Subcar- rier selection for efficient CSI-based indoor localization,”IOP Conf. Ser. Mater. Sci. Eng., vol. 383, no. 1, p. 012017, Jul. 2018

    2018

  49. [49]

    AdaMAE: Adaptive masking for efficient spa- tiotemporal learning with masked autoencoders,

    W. G. C. Bandara, N. Patel, A. Gholami, M. Nikkhah, M. Agrawal, and V . M. Patel, “AdaMAE: Adaptive masking for efficient spa- tiotemporal learning with masked autoencoders,” 2022, arXiv preprint arXiv:2211.09120

    Pith/arXiv arXiv 2022

  50. [50]

    On Learning Intrinsic Rewards for Policy Gradient Methods,

    Z. Zheng, J. Oh, and S. Singh, “On Learning Intrinsic Rewards for Policy Gradient Methods,” 2018, arXiv preprint arXiv:1804.06459

    Pith/arXiv arXiv 2018

  51. [51]

    DISK: Learning local features with policy gradient,

    M. J. Tyszkiewicz, P. Fua, and E. Trulls, “DISK: Learning local features with policy gradient,” 2020, arXiv preprint arXiv:2006.13566

    Pith/arXiv arXiv 2020

  52. [52]

    SenseFi: A library and benchmark on deep-learning-empowered WiFi human sensing,

    J. Yang, X. Chen, D. Wang, H. Zou, C. X. Lu, S. Sun, and L. Xie, “SenseFi: A library and benchmark on deep-learning-empowered WiFi human sensing,” 2023, arXiv preprint arXiv:2207.07859

    Pith/arXiv arXiv 2023

  53. [53]

    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,” 2021, arXiv preprint arXiv:2010.11929

    Pith/arXiv arXiv 2021

  54. [54]

    A dataset for Wi-Fi-based human activity recognition in line-of-sight and non- line-of-sight indoor environments,

    B. A. Alsaify, M. M. Almazari, R. Alazrai, and M. I. Daoud, “A dataset for Wi-Fi-based human activity recognition in line-of-sight and non- line-of-sight indoor environments,”Data in Brief, vol. 33, p. 106534, 2020

    2020

  55. [55]

    SignFi: Sign language recognition using WiFi,

    Y . Ma, G. Zhou, S. Wang, H. Zhao, and W. Jung, “SignFi: Sign language recognition using WiFi,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 2, no. 1, p. 23, Mar. 2018

    2018

  56. [56]

    A Survey on Behavior Recognition Using WiFi Channel State Information,

    S. Yousefi, H. Narui, S. Dayal, S. Ermon, and S. Valaee, “A Survey on Behavior Recognition Using WiFi Channel State Information,”IEEE Commun. Mag., vol. 55, no. 10, pp. 98–104, Oct. 2017

    2017

  57. [57]

    WiMANS: A benchmark dataset for WiFi-based multi-user activity sensing,

    S. Huang, K. Li, D. You, Y . Chen, A. Lin, S. Liu, X. Li, and J. A. McCann, “WiMANS: A benchmark dataset for WiFi-based multi-user activity sensing,” 2024, arXiv preprint arXiv:2402.09430

    Pith/arXiv arXiv 2024

  58. [58]

    Widar 3.0: WiFi-based activity recognition dataset,

    Z. Yang, Y . Zhang, G. Zhang, Y . Zheng, and G. Chi, “Widar 3.0: WiFi-based activity recognition dataset,” IEEE Dataport, 2020

    2020

  59. [59]

    A Survey of Machine Learning in Edge Computing: Techniques, Frameworks, Applications, Issues, and Research Direc- tions,

    O. Jouini, K. Sethom, A. Namoun, N. Aljohani, M. H. Alanazi, and M. N. Alanazi, “A Survey of Machine Learning in Edge Computing: Techniques, Frameworks, Applications, Issues, and Research Direc- tions,”Technologies, vol. 12, no. 6, p. 81, 2024

    2024

  60. [60]

    Tool release: Gathering 802.11n traces with channel state information,

    D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool release: Gathering 802.11n traces with channel state information,”SIGCOMM Comput. Commun. Rev., vol. 41, no. 1, p. 53, Jan. 2011

    2011

  61. [61]

    A CSI-Based Multi-Environment Human Activity Recogni- tion Framework,

    B. A. Alsaify, M. M. Almazari, R. Alazrai, S. Alouneh, and M. I. Daoud, “A CSI-Based Multi-Environment Human Activity Recogni- tion Framework,”Appl. Sci., vol. 12, no. 2, p. 930, 2022

    2022

  62. [62]

    Exploit- ing Wi-Fi Signals for Human Activity Recognition,

    B. A. Alsaify, M. M. Almazari, R. Alazrai, and M. I. Daoud, “Exploit- ing Wi-Fi Signals for Human Activity Recognition,” inProc. 12th Int. Conf. Inf. Commun. Syst. (ICICS), May 2021, pp. 245–250

    2021

  63. [63]

    A dataset for wi-fi-based human-to-human interaction recognition,

    R. Alazrai, A. Awad, B. Alsaify, M. Hababeh, and M. I. Daoud, “A dataset for wi-fi-based human-to-human interaction recognition,”Data in Brief, vol. 31, p. 105668, 2020

    2020

  64. [64]

    You are facing the Mona Lisa: Spot localization using PHY layer information,

    S. Sen, B. Radunovic, R. R. Choudhury, and T. Minka, “You are facing the Mona Lisa: Spot localization using PHY layer information,” in Proc. 10th Int. Conf. Mobile Syst., Appl., Services, 2012, pp. 183– 196

    2012

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.