Pith. sign in

REVIEW 4 major objections 5 minor 88 references

Data Can Speak for Itself: Quality-guided Utilization of Wireless Synthetic Data

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper shows that measuring synthetic wireless data quality with classifier margins, then filtering and pseudo-labeling instead of mixing everything, turns a 13.4 percent degradation into a 4.3 percent gain.

desk verdict A useful, imperfect empirical paper: SynCheck's task-accuracy gains are real, but the quality-metric generality is thinner than the headline suggests. read the letter →

arxiv 2506.23174 v1 pith:FO65E6XM submitted 2025-06-29 cs.LG cs.AI

Chen Gong , Bo Liang , Wei Gao , Chenren Xu This is my paper
classification cs.LGcs.AI
keywords wirelesssensingsyntheticdataqualityaffinitydiversityTRmarginTSsemi-supervisedlearningWiFiCSI
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 argues that the way to make synthetic wireless data useful is not to generate more of it, but to measure its quality and let that measurement decide what enters training. It introduces two margin-based metrics, TR margin and TS margin, which compare classification-confidence distributions of synthetic and real data and can be applied across different datasets. Applied to representative GAN and diffusion wireless generators, the metrics show that current synthetic data are broadly affinity-deficient, so their condition labels are unreliable, while diversity is often preserved. On that diagnosis the paper builds SynCheck, a semi-supervised scheme that filters out low-affinity samples and pseudo-labels the rest; across benchmarks it consistently beats quality-oblivious mixing, reaching a 4.3 percent improvement in the worst case where naive mixing caused a 13.4 percent degradation.

What carries the argument

The load-bearing tool is the classification margin, $z_y(x)-\max_{i\neq y}z_i(x)$, the gap between the model's confidence in the label and its confidence in the strongest alternative class. The paper defines TR margin as the margin of synthetic samples under a model trained on real data, which serves as an affinity probe, and TS margin as the margin of real test samples under a model trained on synthetic data, which serves as a diversity probe. To make these comparisons cross-dataset, margins are calibrated by shifting the test margin distribution by the mean difference between training margins and an in-distribution standard test set, and the remaining gap with the training margin distribution is measured by Jensen-Shannon divergence. SynCheck is the second mechanism: a task classifier plus per-class one-vs-all inlier/outlier detectors, trained first with real labels and consistency regularization on unlabeled synthetic data, then refined by assigning pseudo-labels only to samples the detectors accept, under wireless-specific augmentations.

What would settle it

Re-run the affinity and diversity assessment with a non-convolutional task model, such as an MLP or a small transformer, on the same CsiGAN and RF-Diffusion datasets, and recompute the calibrated TR and TS margin divergences. The central claim is falsified if the quality ordering between the two generators reverses or if SynCheck's reported divergence reductions of 57.3 and 29.5 percent do not appear, because that would show the margins measure the classifier rather than the data.

Watch

Extended reading notes

Core claim

The central claim is that a data quality metric grounded in classifier margins can both diagnose and fix the failure of synthetic data in wireless sensing. Concretely, the paper claims that existing wireless synthetic data suffer from a prevalent affinity limitation: the conditional sample distribution $p_{\theta}(x|y)$ does not match the real $p(x|y)$, which makes condition labels unreliable, while diversity, captured by $p_{\theta}(y|x)\approx p(y|x)$, is largely preserved for in-domain generation. The proposed TR margin, which is the margin of synthetic samples under a model trained on real data, and TS margin, which is the margin of real test samples under a model trained on synthetic data, are measured as Jensen-Shannon divergence between calibrated margin distributions and the training margin distribution. SynCheck treats synthetic data as unlabeled, filters samples its per-class inlier detectors deem low-affinity, and pseudo-labels the rest; the paper reports that this reduces TR-margin divergence by 57.3 percent for CsiGAN and 29.5 percent for RF-Diffusion while leaving diversity nearly unchanged, and that task accuracy improves by about 8.6 percent and 7.2 percent relative to nonselective mixing in the two setups.

Load-bearing premise

The result depends on the calibrated margin gap being a property of the synthetic data rather than of the particular classifier, even though about 5.9 percent variation across architectures and training epochs remains after calibration.

Editorial extensions

If this is right

  • Nonselective mixing of synthetic and real data should be treated as risky: in the in-domain setup, raising synthetic volume to 500 percent of real data turned a 72.5 percent real-only baseline into a 13.4 percent degradation, while SynCheck turned the same data into a 4.3 percent gain.
  • Affinity is the binding constraint for current wireless synthetic data: adding more diverse but low-affinity samples shifts the training distribution and can hurt the task even when diversity is good.
  • The TR and TS margin metrics give wireless generative models a cross-dataset yardstick, so quality can be compared without requiring a shared benchmark dataset.
  • Generation conditions remain valuable even when their labels are not trusted: conditional models produced higher-quality synthetic data than unconditional counterparts, and SynCheck performed worse when it reused generation conditions as labels instead of assigning pseudo-labels.
  • The utilization scheme adds negligible parameters and FLOPs, roughly doubles training GPU memory, and leaves inference cost unchanged, making it practical as a post-processing step.

Reading between the lines

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

  • The same recipe, treating synthetic samples as unlabeled, keeping only detector-confirmed inliers, and pseudo-labeling them under domain-specific augmentation, should transfer to other sensing modalities such as radar, mmWave, or biosignals, because it requires only a classifier and per-class detectors, not human-interpretable signals.
  • Because the margin divergences are computable during training, they could be used online to pause generation or adapt sampling when affinity drops, rather than only filtering after a batch has been produced.
  • The affinity-diversity tradeoff reported here suggests a tunable selection threshold: instead of the fixed inlier/outlier split, one could optimize the filtering boundary directly against downstream task accuracy.
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

4 major / 5 minor

Summary. The paper addresses the question of how to assess and utilize the quality of synthetic data in wireless sensing. It proposes two margin-based quality metrics, TR-margin and TS-margin, defined as the difference between the true-class confidence and the maximum other-class confidence, and measures affinity and diversity via the Jensen-Shannon divergence between the margin distributions of a reference training set and a test set, after a mean-based calibration step. Using these metrics on two open-source wireless generative models (CsiGAN for cross-domain transfer and RF-Diffusion for in-domain amplification), the paper reports a prevalent affinity limitation in current wireless synthetic data. It then introduces SynCheck, a semi-supervised framework that treats synthetic data as unlabeled, filters low-affinity samples via per-class inlier/outlier detectors, assigns pseudo-labels to the remaining samples, and trains with consistency and entropy losses. Evaluations on SignFi and Widar show that SynCheck improves task accuracy over real-data-only training and over nonselective mixing of real and synthetic data, with the largest gains when the nonselective baseline degrades performance.

Significance. If the proposed metrics are indeed generalizable, they would give the wireless community a cheap, task-agnostic way to decide whether and how to use synthetic data, a problem of clear practical importance. SynCheck is a sensible post-processing approach that does not require retraining generative models. The paper also makes a useful distinction between cross-domain and in-domain generation and connects quality deficiencies to concrete causes (untrained conditions, domain-specific processing). However, the strength of these contributions is currently weakened by evaluation gaps: the generalizability of the margin-based metrics is not fully established, the quality re-evaluation in Section 7.2 risks circularity, and the central performance comparisons lack error bars and significance information. The core ideas are sound enough to warrant a revised version.

major comments (4)
  1. [§4.2 (Margin Calibration) and RQ1] The claim that the calibrated TR/TS margin JS divergence is a model- and dataset-independent quality measure is not established. The necessity-of-calibration analysis reports uncalibrated variation of 13.5% and calibrated variation of 5.9% across ResNet18/34 and training epochs, but calibration is only a mean shift and cannot remove architecture-dependent differences in the shape of the margin distribution, which is exactly what JS divergence measures. The RQ1 verdict of "prevalent limited affinity" is based on margins from a ResNet34 model on two datasets. To make this diagnosis load-bearing, please evaluate the calibrated margin distributions with at least one non-ResNet task model family (e.g., a simple MLP, a transformer, or a lightweight CNN) and report whether the CsiGAN-vs-RF-Diffusion ordering and the limited-affinity verdict are preserved. If they flip, the metrics are not generalizable and the motivation for SynCheck becomes unsupported.
  2. [§7.2 (Quality Re-evaluation) and Table 3] The re-evaluation of affinity and diversity after SynCheck filtering does not state whether the TR/TS margins are computed with a model trained only on real data or with the SynCheck-trained model that performed filtering and pseudo-label assignment. If the latter, the reported 57.3% and 29.5% reductions in TR-margin JS divergence are mechanically biased: a model that selects samples by its own confidence and is then trained on those samples will naturally produce high confidence margins on them, so the improvement may reflect selection-induced overconfidence rather than intrinsic affinity. Please clarify the provenance of the margins, and if the SynCheck model was used, re-run the evaluation with a model trained only on real data (or an independent held-out model) and report the results. Without this, the claimed alignment between task performance and quality metrics is circular.
  3. [§7.1.3 and Figures 9–13, Table 4] All reported task accuracies are single scalar values without standard deviations, confidence intervals, or counts of random seeds. The abstract's claim that SynCheck "consistently outperforms" quality-oblivious utilization requires evidence that the gaps are not within run-to-run variation. Please report mean and standard deviation over at least 5 independent seeds for the key comparisons in Fig. 9, Fig. 11, Fig. 12, and Table 4, and where feasible include paired significance tests (e.g., paired bootstrap or paired t-test). This is particularly important because some reported differences, such as SynCheck vs. TRTS filtering (82.7 vs. 81.0 in cross-domain; 75.6 vs. 73.0 in in-domain), are modest and could be within noise.
  4. [§4.1 and Abstract] The conclusion that "current wireless synthetic data" suffer from "prevalent affinity limitation" is drawn from only two open-source models, one per taxonomy category. This sample is too narrow to support a general prevalence claim. Please either restrict the wording to the evaluated models (e.g., "the representative open-source models we evaluated") or include additional representative models from the closed-source literature, even with approximate reproduction, to substantiate the prevalence claim. As written, the abstract and introduction overgeneralize from two model-dataset pairs.
minor comments (5)
  1. [§3.2 (Metric Advantage)] The Pearson correlation analysis does not state the number of data points used for each correlation. With Gaussian noise varied over 6 steps (0.05 to 0.3), the affinity correlation is based on only 6 points; with 10 steps for diversity, the analysis is still small-sample. Please report n and consider permutation-based p-values or confidence intervals for the correlations.
  2. [Table 1] The CAS row contains a "( missing)" placeholder in the Quality Metrics column, which appears to be an incomplete draft artifact. Please fill in the entry or remove the placeholder.
  3. [§4.2 (Abnormal Train Margins, Fig. 5d)] The observation that roughly 20% of train margins are negative in the train-synthetic setup would be more convincing with a quantitative statement of the fraction and a sanity check that a model trained on real data does not show a similar pattern. As presented, the observation is based on visual inspection of a histogram.
  4. [§7.4 (Impact of Synthetic Data Amount)] The sentence "with 50% synthetic data, SynCheck achieves performance metrics of 81.8% and 75.5% for in-domain and cross-domain setups, respectively" appears to swap the domain order relative to Fig. 12 and to the values reported for the nonselective mixture (76.8% and 70.5%). Please verify and correct the domain labels in the text.
  5. [§5.2 (Self-supervision with Unlabeled Synthetic Data)] The consistency loss uses Gaussian noise as the perturbation for wireless signals, but the text does not justify why additive Gaussian noise is an appropriate perturbation for CSI/DFS inputs. A brief justification or citation (beyond referring to propagation channel noise) would help the reader assess the choice.

Circularity Check

1 steps flagged · score 4.0 of 10

The headline task-performance comparison is externally benchmarked and not circular, but the Section 7.2 quality re-evaluation of 'enhanced affinity' is partly self-fulfilling because the same SynCheck-trained model that was fit on the filtered pseudo-labeled samples is used to compute the TR margins.

  1. fitted input called prediction [Section 7.2 (Quality Re-evaluation), with Eq. (4) in Section 5.3 and Definition 3 in Section 3.1]
    "We re-evaluate the quality attributes of affinity and diversity for the filtered and relabeled synthetic data with SynCheck, as shown by the 'test' in Fig. 10. ... Our results show that SynCheck significantly improves the TR-margin, achieving a 57.3% and 29.5% reduction in JS Divergence for TR margins, thereby enhancing data affinity."

    The TR margin (Definition 3) is the classification confidence difference of the task model. The re-evaluation is described as being performed 'with SynCheck,' and no separate real-only model is specified, so the margins are naturally those of the SynCheck-trained classifier. That classifier was trained in Phase 2 on exactly the filtered synthetic samples with Eq. (4), Lpseu = cross entropy(A_s(s_b), y_hat), which directly maximizes the model's confidence for the assigned pseudo-labels on the selected samples. Measuring the TR margin on those same samples with that same model does not provide independent evidence of improved affinity; the reported 57.3%/29.5% JS-divergence reduction is partly the training objective evaluated on its own training set.

full rationale

Verdict: the paper's central task-performance result is not circular. SynCheck is benchmarked against real-data-only, nonselective mixture, SSIM filtering, and TRTS filtering, using held-out classification accuracy as an external metric; none of these comparisons reduces to the paper's own fitted values. The Bayesian Theorems 1 and 2 and the margin-based Conclusions 1 and 2 are derived from stated assumptions and cite external margin-generalization results; no load-bearing self-citation chain exists. The one partial circularity is in Section 7.2's quality re-evaluation: 'filtered and relabeled synthetic data with SynCheck' are scored with the TR margin, which is the confidence difference of a classifier that has just been trained with Eq. (4) to output the assigned pseudo-labels on exactly those filtered samples. The reported 57.3% and 29.5% JS-divergence reduction in TR margin is therefore at least partly a self-fulfilling consequence of the training objective rather than an independent confirmation that the selected synthetic data have higher intrinsic affinity. A separate validity concern, not a circularity, is the residual 5.9% variation in calibrated margins across architectures and epochs reported in Section 4.2, which leaves the RQ1 'prevalent limited affinity' diagnosis somewhat model-dependent. Overall, the main performance claims stand on independent ground, so the circularity score is moderate rather than high.

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

The central claims rest on standard probability bounds and on the domain assumption that training-set margin distributions, after mean-shift calibration, are a reliable cross-dataset reference for synthetic data quality. The free parameters listed are design choices in the calibration and regularization steps that are not reported numerically. No new physical entities are introduced.

free parameters (2)
  • Calibration standard-test subset size
    The fraction of the training set held out as the standard test set in the margin calibration procedure is not reported; it is a hand-chosen quantity that affects the calibration shift.
  • Consistency noise level
    The Gaussian noise standard deviation used for consistency regularization in SynCheck's phase 1 is not specified; it is a design choice that affects pseudo-label quality.
assumptions (4)
  • standard math Total variation distance bounds test loss via conditional distribution mismatch
    Used in the proofs of Theorems 1 and 2 in Appendix A to connect affinity and diversity to task performance.
  • domain assumption The margin distribution of a model on its training set is a reliable reference for generalization
    Invoked in Section 3.1 to justify using training-set margins as a standard reference for cross-dataset quality comparison.
  • domain assumption A task model trained on real data provides a meaningful reference signal for evaluating synthetic data affinity
    Needed so that TR margins reflect synthetic data quality independent of the generative model.
  • domain assumption Synthetic data quality can be decomposed into affinity and diversity as defined in Definitions 1 and 2
    This decomposition inherits from prior work and is not independently proven for wireless data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Data Can Speak for Itself: Quality-guided Utilization of Wireless Synthetic Data." pith.science (2026). https://pith.science/paper/FO65E6XM

@misc{pith2026250623174,
  author       = {Pith},
  title        = {Pith review of: Data Can Speak for Itself: Quality-guided Utilization of Wireless Synthetic Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FO65E6XM}},
  note         = {Machine review of arXiv:2506.23174}
}
read the original abstract

Generative models have gained significant attention for their ability to produce realistic synthetic data that supplements the quantity of real-world datasets. While recent studies show performance improvements in wireless sensing tasks by incorporating all synthetic data into training sets, the quality of synthetic data remains unpredictable and the resulting performance gains are not guaranteed. To address this gap, we propose tractable and generalizable metrics to quantify quality attributes of synthetic data - affinity and diversity. Our assessment reveals prevalent affinity limitation in current wireless synthetic data, leading to mislabeled data and degraded task performance. We attribute the quality limitation to generative models' lack of awareness of untrained conditions and domain-specific processing. To mitigate these issues, we introduce SynCheck, a quality-guided synthetic data utilization scheme that refines synthetic data quality during task model training. Our evaluation demonstrates that SynCheck consistently outperforms quality-oblivious utilization of synthetic data, and achieves 4.3% performance improvement even when the previous utilization degrades performance by 13.4%.

Figures

Figures reproduced from arXiv: 2506.23174 by the authors.

Figure 1
Figure 1. The labeled synthetic data has proved effective to improve the performance of data-hungry wireless applications such as action recognition [11] and indoor localization [12]. While most work in data synthesis for data scarcity alleviation approves its success in supplementing quantity [13–15], prior stud￾ies have highlighted a remaining deficiency of synthetic data quality in capturing domain-specific knowledge, such… view at source ↗
Figure 1
Figure 1. Our quality assessment and quality-guided utilization of wireless synthetic data. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Our tractable and generalizable metrics: TR margin [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Calibration of the margin. the SignFi dataset [58] for CsiGAN[6], and the Widar dataset [59] for RF-Diffusion [9]. Margin Calibration. Our metrics employ margin comparisons between the test set and the train set to enable generalizable quality quantification across dat…
Figure 6
Figure 6. Figure 6: Diversity comparison between trained and un [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 8
Figure 8. Figure 8: SynCheck Workflow. (outliers) and assign pseudo-labels to the remaining high-affinity synthetic data (inliers). The pseudo-label supervision iteratively improves the model’s performance in a domain-dependent manner. Pseudo-label Assignment. For each synthetic sample 𝑢,…
Figure 9
Figure 9. Figure 9: Overall Performance. RQ4: How does SynCheck perform in task performance and data quality attributes of affinity and diversity? RQ5: What is the impact of design choices, including loss com￾ponents, backbone model, synthetic data volume, and label origin, in the SynChec…
Figure 10
Figure 10. Figure 10: Quality re-quantification of the filtered synthetic [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 11
Figure 11. Figure 11: Performance v.s. Backbone Models. SynCheck. In our semi-supervised learning framework, the back￾bone model functions as a feature extractor, providing outputs that are used by task classifier 𝑓 and per-class inlier-outlier detector 𝑔 𝑗 . Impact of Backbone Model. The …
Figure 13
Figure 13. Figure 13: Condition assessment: pseudo-label origin and [PITH_FULL_IMAGE:figures/full_fig_p011_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

88 extracted references · 71 canonical work pages

  1. [1]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020

    2020

  2. [2]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. NeurIPS, 2014

    2014

  3. [3]

    Generating diverse high- fidelity images with vq-vae-2

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high- fidelity images with vq-vae-2. NeurIPS, 2019

    2019

  4. [4]

    Conditional generative adversarial nets

    Mehdi Mirza and Simon Osindero. Conditional generative adversarial nets. ArXiv:1411.1784, 2014

    arXiv 2014

  5. [5]

    Ganwriting: Content-conditioned generation of styled handwritten word images

    Lei Kang, Pau Riba, Yaxing Wang, Marçal Rusiñol, Alicia Forn’es, and Mauricio Villegas. Ganwriting: Content-conditioned generation of styled handwritten word images. Springer ECCV, 2020

    2020

  6. [6]

    Csigan: Robust channel state information-based activity recognition with gans

    Chunjing Xiao, Daojun Han, Yongsen Ma, and Zhiguang Qin. Csigan: Robust channel state information-based activity recognition with gans. IEEE Internet of Things Journal, 2019

    2019

  7. [7]

    Cross-frequency training with adversarial learning for radar micro-doppler signature classification

    Sevgi Zubeyde Gurbuz, Mohammad Muntasir Rahman, Emre Kurtoğlu, Trevor Macks, and Francesco Fioranelli. Cross-frequency training with adversarial learning for radar micro-doppler signature classification. In Commercial Sensing, 2020

    2020

  8. [8]

    Fido: Ubiquitous fine-grained wifi-based localization for unlabelled users via domain adaptation

    Xi Chen, Hang Li, Chenyi Zhou, Xue Liu, Di Wu, and Gregory Dudek. Fido: Ubiquitous fine-grained wifi-based localization for unlabelled users via domain adaptation. In ACM WWW, 2020

    2020

Show all 88 references
  1. [9]

    Rf-diffusion: Radio signal generation via time-frequency diffusion

    Guoxuan Chi, Zheng Yang, Chenshu Wu, Jingao Xu, Yuchong Gao, Yunhao Liu, and Tony Xiao Han. Rf-diffusion: Radio signal generation via time-frequency diffusion. In ACM MobiCom, 2024

    2024

  2. [10]

    Rf genesis: Zero-shot generalization of mmwave sensing through simulation-based data synthesis and generative diffusion models

    Xingyu Chen and Xinyu Zhang. Rf genesis: Zero-shot generalization of mmwave sensing through simulation-based data synthesis and generative diffusion models. In ACM SenSys, 2023

    2023

  3. [11]

    Crossgr: Accurate and low-cost cross-target gesture recognition using wi-fi

    Xinyi Li, Liqiong Chang, Fangfang Song, Ju Wang, Xiaojiang Chen, Zhanyong Tang, and Zheng Wang. Crossgr: Accurate and low-cost cross-target gesture recognition using wi-fi. ACM IMWUT, 2021

    2021

  4. [12]

    Fidora: Robust wifi-based indoor localization via unsupervised domain adaptation

    Xi Chen, Hang Li, Chenyi Zhou, Xue Liu, Di Wu, and Gregory Dudek. Fidora: Robust wifi-based indoor localization via unsupervised domain adaptation. IEEE Internet of Things Journal , 2022

    2022

  5. [13]

    Medical image generation using generative adversarial networks: A review

    Nripendra Kumar Singh and Khalid Raza. Medical image generation using generative adversarial networks: A review. Health informatics: A computational perspective in healthcare, 2021

    2021

  6. [14]

    Cross-domain wifi sensing with channel state information: A survey

    Chen Chen, Gang Zhou, and Youfang Lin. Cross-domain wifi sensing with channel state information: A survey. ACM Computing Surveys, 2023

    2023

  7. [15]

    Learning to sense: Deep learning for wireless sensing with less training efforts

    Jie Wang, Qinhua Gao, Xiaorui Ma, Yunong Zhao, and Yuguang Fang. Learning to sense: Deep learning for wireless sensing with less training efforts. IEEE Wireless Communications, 2020

    2020

  8. [16]

    What does dall-e 2 know about radiology? Journal of Medical Internet Research, 2023

    Lisa C Adams, Felix Busch, Daniel Truhn, Marcus R Makowski, Hugo JWL Aerts, and Keno K Bressem. What does dall-e 2 know about radiology? Journal of Medical Internet Research, 2023

    2023

  9. [17]

    Adapting pretrained vision-language foundational models to medical imaging domains

    Pierre Chambon, Christian Bluethgen, Curtis P Langlotz, and Akshay Chaudhari. Adapting pretrained vision-language foundational models to medical imaging domains. arXiv:2210.04133, 2022

    2022 arXiv

  10. [18]

    Aligning synthetic medical images with clinical knowledge using human feedback

    Shenghuan Sun, Greg Goldgof, Atul Butte, and Ahmed M Alaa. Aligning synthetic medical images with clinical knowledge using human feedback. NeurIPS, 2024

    2024

  11. [19]

    How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models

    Ahmed Alaa, Boris Van Breugel, Evgeny S Saveliev, and Mihaela van der Schaar. How faithful is your synthetic data? sample-level metrics for evaluating and auditing generative models. In ICML, 2022

    2022

  12. [20]

    Synthetic data in machine learning for medicine and healthcare

    Richard J Chen, Ming Y Lu, Tiffany Y Chen, Drew FK Williamson, and Faisal Mahmood. Synthetic data in machine learning for medicine and healthcare. Nature Biomedical Engineering, 2021

    2021

  13. [21]

    Are gans created equal? a large-scale study

    Mario Lucic, Karol Kurach, Marcin Michalski, Sylvain Gelly, and Olivier Bousquet. Are gans created equal? a large-scale study. NeurIPS, 2018

    2018

  14. [22]

    Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models

    George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecroze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza- Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS, 2024

    2024

  15. [23]

    Uwb-fi: Pushing wi-fi towards ultra-wideband for fine-granularity sensing

    Xin Li, Hongbo Wang, Zhe Chen, Zhiping Jiang, and Jun Luo. Uwb-fi: Pushing wi-fi towards ultra-wideband for fine-granularity sensing. In MobiSys, 2024

    2024

  16. [24]

    Mmbind: Unleashing the potential of distributed and heterogeneous data for multimodal learning in iot

    Xiaomin Ouyang, Jason Wu, Tomoyoshi Kimura, Yihan Lin, Gunjan Verma, Tarek Abdelzaher, and Mani Srivastava. Mmbind: Unleashing the potential of distributed and heterogeneous data for multimodal learning in iot. arXiv:2411.12126, 2024

    2024 arXiv

  17. [25]

    Llmsense: Harnessing llms for high-level reasoning over spatiotemporal sensor traces

    Xiaomin Ouyang and Mani Srivastava. Llmsense: Harnessing llms for high-level reasoning over spatiotemporal sensor traces. arXiv:2403.19857, 2024

    2024 arXiv

  18. [26]

    Advancing multi-modal sensing through expandable modality alignment

    Shenghong Dai, Shiqi Jiang, Yifan Yang, Ting Cao, Mo Li, Suman Banerjee, and Lili Qiu. Advancing multi-modal sensing through expandable modality alignment. arXiv:2407.17777, 2024

    2024 arXiv

  19. [27]

    How good is my gan? In Springer ECCV, 2018

    Konstantin Shmelkov, Cordelia Schmid, and Karteek Alahari. How good is my gan? In Springer ECCV, 2018

    2018

  20. [28]

    Spectrally-normalized margin bounds for neural networks

    Peter L Bartlett, Dylan J Foster, and Matus J Telgarsky. Spectrally-normalized margin bounds for neural networks. NeurIPS, 2017

    2017

  21. [29]

    Large margin deep networks for classification

    Gamaleldin Elsayed, Dilip Krishnan, Hossein Mobahi, Kevin Regan, and Samy Bengio. Large margin deep networks for classification. NeurIPS, 2018

    2018

  22. [30]

    Identifying mislabeled data using the area under the margin ranking

    Geoff Pleiss, Tianyi Zhang, Ethan Elenberg, and Kilian Q Weinberger. Identifying mislabeled data using the area under the margin ranking. NeurIPS, 2020

    2020

  23. [31]

    Understanding deep learning (still) requires rethinking generalization

    Chiyuan Zhang, Samy Bengio, Moritz Hardt, Benjamin Recht, and Oriol Vinyals. Understanding deep learning (still) requires rethinking generalization. Commu- nications of the ACM, 2021

    2021

  24. [32]

    Towards generalized mmwave-based human pose estimation through signal augmentation

    Hongfei Xue, Qiming Cao, Chenglin Miao, Yan Ju, Haochen Hu, Aidong Zhang, and Lu Su. Towards generalized mmwave-based human pose estimation through signal augmentation. In ACM MobiCom, 2023

    2023

  25. [33]

    Quality aware generative adversarial networks

    Kancharla Parimala and Sumohana Channappayya. Quality aware generative adversarial networks. NeurIPS, 32, 2019

    2019

  26. [34]

    Classification accuracy score for conditional generative models

    Suman Ravuri and Oriol Vinyals. Classification accuracy score for conditional generative models. NeurIPS, 2019

    2019

  27. [35]

    Teaching rf to sense without rf training measurements

    Hong Cai, Belal Korany, Chitra R Karanam, and Yasamin Mostofi. Teaching rf to sense without rf training measurements. ACM IMWUT, 2020

    2020

  28. [36]

    Wifi sensing with channel state information: A survey

    Yongsen Ma, Gang Zhou, and Shuangquan Wang. Wifi sensing with channel state information: A survey. ACM Computing Surveys, 2019

    2019

  29. [37]

    Survey of time series data generation in iot

    Chaochen Hu, Zihan Sun, Chao Li, Yong Zhang, and Chunxiao Xing. Survey of time series data generation in iot. Sensors, 2023

    2023

  30. [38]

    Rfboost: Understanding and boosting deep wifi sensing via physical data augmentation

    Weiying Hou and Chenshu Wu. Rfboost: Understanding and boosting deep wifi sensing via physical data augmentation. ACM IMWUT, 2024

    2024

  31. [39]

    Simple and effective augmentation methods for csi based indoor localization

    Omer Gokalp Serbetci, Ju-Hyung Lee, Daoud Burghal, and Andreas F Molisch. Simple and effective augmentation methods for csi based indoor localization. In IEEE GLOBECOM, 2023

    2023

  32. [40]

    Data augmentation techniques for cross- domain wifi csi-based human activity recognition

    Julian Strohmayer and Martin Kampel. Data augmentation techniques for cross- domain wifi csi-based human activity recognition. In IFIP AIAI, 2024

    2024

  33. [41]

    Ray tracing as a design tool for radio networks

    John W McKown and R Lee Hamilton. Ray tracing as a design tool for radio networks. IEEE Network, 1991

    1991

  34. [42]

    Nerf2: Neural radio- frequency radiance fields

    Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. Nerf2: Neural radio- frequency radiance fields. In ACM MobiCom, 2023

    2023

  35. [43]

    Food and liquid sensing in practical environments using{RFIDs}

    Unsoo Ha, Junshan Leng, Alaa Khaddaj, and Fadel Adib. Food and liquid sensing in practical environments using{RFIDs}. In NSDI, 2020

    2020

  36. [44]

    Survey on synthetic data generation, evaluation methods and gans

    Alvaro Figueira and Bruno Vaz. Survey on synthetic data generation, evaluation methods and gans. MDPI Mathematics, 2022

    2022

  37. [45]

    Pros and cons of gan evaluation measures

    Ali Borji. Pros and cons of gan evaluation measures. Elsevier Computer vision and image understanding, 2019

    2019

  38. [46]

    Quality aware generative adversarial networks

    Kancharla Parimala and Sumohana Channappayya. Quality aware generative adversarial networks. NeurIPS, 2019

    2019

  39. [47]

    Gans trained by a two time-scale update rule converge to a local nash equilibrium

    Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017

    2017

  40. [48]

    Categorical generative model evaluation via synthetic distribution coarsening

    Florence Regol and Mark Coates. Categorical generative model evaluation via synthetic distribution coarsening. In AISTATS, 2024

    2024

  41. [49]

    Bayes’ theorem — Wikipedia, the free encyclopedia,

    Wikipedia contributors. Bayes’ theorem — Wikipedia, the free encyclopedia,

  42. [50]

    Medical image synthesis with context-aware generative adversarial networks

    Dong Nie, Roger Trullo, Jun Lian, Caroline Petitjean, Su Ruan, Qian Wang, and Dinggang Shen. Medical image synthesis with context-aware generative adversarial networks. In Springer MICCAI, 2017

    2017

  43. [51]

    Cross-scenario device-free activity recognition based on deep adversarial networks

    Jie Wang, Yunong Zhao, Xiaorui Ma, Qinghua Gao, Miao Pan, and Hongyu Wang. Cross-scenario device-free activity recognition based on deep adversarial networks. IEEE Transactions on Vehicular Technology, 2020

    2020

  44. [52]

    A deep-learning-based self-calibration time-reversal fingerprinting localization approach on wi-fi plat- form

    Lili Zheng, Bin-Jie Hu, Jinguang Qiu, and Manman Cui. A deep-learning-based self-calibration time-reversal fingerprinting localization approach on wi-fi plat- form. IEEE Internet of Things Journal , 2020

    2020

  45. [53]

    Taming the inconsistency of wi-fi fingerprints for device-free passive indoor localization

    Xi Chen, Chen Ma, Michel Allegue, and Xue Liu. Taming the inconsistency of wi-fi fingerprints for device-free passive indoor localization. In IEEE INFOCOM, 2017

    2017

  46. [54]

    Unsupervised and semi-supervised learning with categorical generative adversarial networks

    Jost Tobias Springenberg. Unsupervised and semi-supervised learning with categorical generative adversarial networks. In ICLR, 2016

    2016

  47. [55]

    Af-dcgan: Amplitude feature deep convolutional gan for fingerprint construc- tion in indoor localization systems

    Qiyue Li, Heng Qu, Zhi Liu, Nana Zhou, Wei Sun, Stephan Sigg, and Jie Li. Af-dcgan: Amplitude feature deep convolutional gan for fingerprint construc- tion in indoor localization systems. IEEE Transactions on Emerging Topics in Computational Intelligence, 2021

    2021

  48. [56]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE/CVF CVPR, 2016

    2016

  49. [57]

    Sensefi: A library and benchmark on deep-learning- empowered wifi human sensing

    Jianfei Yang, Xinyan Chen, Dazhuo Wang, Han Zou, Chris Xiaoxuan Lu, Sumei Sun, and Lihua Xie. Sensefi: A library and benchmark on deep-learning- empowered wifi human sensing. Elsevier Patterns, 2023

    2023

  50. [58]

    Signfi: Sign language recognition using wifi

    Yongsen Ma, Gang Zhou, Shuangquan Wang, Hongyang Zhao, and Woosub Jung. Signfi: Sign language recognition using wifi. ACM IWMUT, 2018

    2018

  51. [59]

    Zero-effort cross-domain gesture recognition with wi-fi

    Yue Zheng, Yi Zhang, Kun Qian, Guidong Zhang, Yunhao Liu, Chenshu Wu, and Zheng Yang. Zero-effort cross-domain gesture recognition with wi-fi. InMobiSys, 2019

    2019

  52. [60]

    On large-batch training for deep learning: Generaliza- tion gap and sharp minima

    Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generaliza- tion gap and sharp minima. arXiv:1609.04836, 2016

    2016 arXiv

  53. [61]

    Sensitivity and generalization in neural networks: an empirical study

    Roman Novak, Yasaman Bahri, Daniel A Abolafia, Jeffrey Pennington, and Jascha Sohl-Dickstein. Sensitivity and generalization in neural networks: an empirical study. arXiv:1802.08760, 2018

    2018 arXiv

  54. [62]

    The jensen-shannon divergence

    María Luisa Menéndez, JA Pardo, L Pardo, and MC Pardo. The jensen-shannon divergence. Journal of the Franklin Institute , 1997

    1997

  55. [63]

    Position and orientation agnostic gesture recognition using wifi

    Aditya Virmani and Muhammad Shahzad. Position and orientation agnostic gesture recognition using wifi. In MobiSys, 2017

    2017

  56. [64]

    Diffar: Adaptive condi- tional diffusion model for temporal-augmented human activity recognition

    Shuokang Huang, Po-Yu Chen, and Julie Ann McCann. Diffar: Adaptive condi- tional diffusion model for temporal-augmented human activity recognition. In IJCAI, 2023

    2023

  57. [65]

    Opencos: Con- trastive semi-supervised learning for handling open-set unlabeled data

    Jongjin Park, Sukmin Yun, Jongheon Jeong, and Jinwoo Shin. Opencos: Con- trastive semi-supervised learning for handling open-set unlabeled data. In Springer ECCV, 2022

    2022

  58. [66]

    Iomatch: Simplifying open-set semi-supervised learning with joint inliers and outliers utilization

    Zekun Li, Lei Qi, Yinghuan Shi, and Yang Gao. Iomatch: Simplifying open-set semi-supervised learning with joint inliers and outliers utilization. In IEEE/CVF ICCV, 2023

    2023

  59. [67]

    Ovanet: One-vs-all network for universal domain adaptation

    Kuniaki Saito and Kate Saenko. Ovanet: One-vs-all network for universal domain adaptation. In IEEE/CVF ICCV, 2021

    2021

  60. [68]

    Temporal ensembling for semi-supervised learning

    Samuli Laine and Timo Aila. Temporal ensembling for semi-supervised learning. arXiv:1610.02242, 2016

    2016 arXiv

  61. [69]

    Proakis and M

    J.G. Proakis and M. Salehi. Digital Communications. McGraw-Hill, 2008

    2008

  62. [70]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. NeurIPS, 2020

    2020

  63. [71]

    Pytorch: An imperative style, high-performance deep learning library

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019

    2019

  64. [72]

    Csigan: Robust channel state information-based activity recognition with gans

    Chunjing Xiao, Daojun Han, Yongsen Ma, and Zhiguang Qin. Csigan: Robust channel state information-based activity recognition with gans. https://github.c om/ChunjingXiao/CsiGAN, 2019

    2019

  65. [73]

    Rfboost: Understanding and boosting deep wifi sensing via physical data augmentation

    Weiying Hou and Chenshu Wu. Rfboost: Understanding and boosting deep wifi sensing via physical data augmentation. https://github.com/aiot-lab/RFBoost, 2024

    2024

  66. [74]

    Conditional generation from unconditional diffusion models using denoiser representations

    Alexandros Graikos, Srikar Yellapragada, and Dimitris Samaras. Conditional generation from unconditional diffusion models using denoiser representations. arXiv:2306.01900, 2023

    2023 arXiv

  67. [75]

    Vaes meet diffusion models: Efficient and high-fidelity generation

    Kushagra Pandey, Avideep Mukherjee, Piyush Rai, and Abhishek Kumar. Vaes meet diffusion models: Efficient and high-fidelity generation. InNeurIPS Workshop on Deep Generative Models and Downstream Applications , 2021

    2021

  68. [76]

    Affinity and diversity: Quantifying mechanisms of data augmentation

    Raphael Gontijo-Lopes, Sylvia J Smullin, Ekin D Cubuk, and Ethan Dyer. Affinity and diversity: Quantifying mechanisms of data augmentation. arXiv:2002.08973, 2020

    2002 arXiv

  69. [77]

    Hide-and-seek privacy challenge: Synthetic data generation vs

    James Jordon, Daniel Jarrett, Evgeny Saveliev, Jinsung Yoon, Paul Elbers, Patrick Thoral, Ari Ercole, Cheng Zhang, Danielle Belgrave, and Mihaela van der Schaar. Hide-and-seek privacy challenge: Synthetic data generation vs. patient re-identification. In NeurIPS Competition an...

    2021

  70. [78]

    Flow-gan: Combining maximum likelihood and adversarial learning in generative models

    Aditya Grover, Manik Dhar, and Stefano Ermon. Flow-gan: Combining maximum likelihood and adversarial learning in generative models. In AAAI, 2018

    2018

  71. [79]

    A complete recipe for diffusion generative models

    Kushagra Pandey and Stephan Mandt. A complete recipe for diffusion generative models. In IEEE/CVF ICCV, 2023

    2023

  72. [80]

    Newrf: A deep learning framework for wireless radiation field reconstruction and channel prediction

    Haofan Lu, Christopher Vattheuer, Baharan Mirzasoleiman, and Omid Abari. Newrf: A deep learning framework for wireless radiation field reconstruction and channel prediction. In ICML, 2024

    2024

  73. [81]

    A review on outlier/anomaly detection in time series data

    Ane Blázquez-García, Angel Conde, Usue Mori, and Jose A Lozano. A review on outlier/anomaly detection in time series data. ACM computing surveys, 2021

    2021

  74. [82]

    Deep learning for anomaly detection

    Ruoying Wang, Kexin Nie, Tie Wang, Yang Yang, and Bo Long. Deep learning for anomaly detection. In WSDM, 2020

    2020

  75. [83]

    Model collapse demystified: The case of regression

    Elvis Dohmatob, Yunzhen Feng, and Julia Kempe. Model collapse demystified: The case of regression. In NeurIPS, 2024

    2024

  76. [84]

    A tale of tails: Model collapse as a change of scaling laws

    Elvis Dohmatob, Yunzhen Feng, Pu Yang, Francois Charton, and Julia Kempe. A tale of tails: Model collapse as a change of scaling laws. ICML, 2024

    2024

  77. [85]

    Beyond model collapse: Scaling up with synthesized data requires verification

    Yunzhen Feng, Elvis Dohmatob, Pu Yang, Francois Charton, and Julia Kempe. Beyond model collapse: Scaling up with synthesized data requires verification. In ICLR, 2024

    2024

  78. [86]

    Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks

    Dong-Hyun Lee. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, 2013

    2013

  79. [87]

    Total variation distance of probability measures — Wikipedia, the free encyclopedia, 2025

    Wikipedia contributors. Total variation distance of probability measures — Wikipedia, the free encyclopedia, 2025. [Online; accessed 18-March-2025]

    2025

  80. [2025]

    [Online; accessed 29-March-2025]

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.