Pith. sign in

REVIEW 3 major objections 6 minor 31 references

Lungmix: A Mixup-Based Strategy for Generalization in Respiratory Sound Classification

T0 review · 3 major / 6 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 Lungmix improves respiratory sound classification on unseen datasets by up to 3.55 points by mixing waveforms with loudness masks and OR-combining labels.

desk verdict Lungmix is a plausible mixup variant with a real evaluation flaw: the headline gain is selected on the test set without repeated seeds or a proper four-class metric. read the letter →

arxiv 2501.00064 v1 pith:Q67KU3MQ submitted 2024-12-29 cs.SD cs.LGeess.AS

Shijia Ge , Weixiang Zhang , Shuzhao Xie , Baixu Yan , Zhi Wang This is my paper · ORCID
classification cs.SDcs.LGeess.AS
keywords respiratorysoundclassificationdomaingeneralizationMixupdataaugmentationAudioSpectrogramTransformerloudnessmaskmulti-labelsingle-source
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

Lungmix is a data-augmentation method for a specific failure: respiratory sound classifiers trained on one dataset lose more than 30 points when tested on another dataset collected with different equipment and annotation rules. The paper argues that ordinary Mixup, which linearly interpolates labels and blends whole waveforms, is a poor fit because abnormal sounds such as crackle and wheeze are sparse and their labels do not blend linearly. Lungmix instead mixes waveforms with a loudness-derived mask plus a random mask and combines labels with a bitwise OR, so a normal sound mixed with a crackle is still labeled crackle. Across the ICBHI, SPRSound, and HF Lung V1 datasets, the method raises the average sensitivity/specificity score by up to 3.55 percentage points over the no-Mixup baseline on combined unseen test data.

What carries the argument

The load-bearing machinery is the mask construction and the label rule. The loudness mask $M_i = |x_i| > |\operatorname{mean}(x_i) + 2\operatorname{std}(x_i)|$ flags the unusual, information-rich parts of each waveform; OR-ing the two loudness masks, scaling by $\lambda \sim \mathrm{Beta}(\alpha,\alpha)$, and OR-ing with a random binary mask $R$ yields samples that keep salient events from both inputs while remaining plausible waveforms. The label rule is bitwise OR over one-hot abnormal-class vectors: normal plus crackle stays crackle, and crackle plus wheeze becomes both, with the combination mapped through Label Powerset to a class. This is what carries the generalization claim: the augmented data inherit the exact acoustic signature and label semantics of the abnormal sounds, rather than a blurred average of two different conditions.

What would settle it

Train an Audio Spectrogram Transformer on HF with Lungmix, then evaluate on ICBHI and SPR test sets and report the full four-class confusion matrix. If the 3.55-point gain comes mostly from higher specificity while per-class sensitivity for crackle and wheeze is flat or worse than the no-Mixup baseline, the claim that Lungmix improves abnormal-sound generalization would be undercut.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that Mixup works for respiratory sounds only when the sparse abnormal events survive the mixture and the label keeps its semantic meaning. Standard Mixup assumes a linear relationship between the input mixture and the label, and the paper shows this assumption hurts transformer-based classification; Patch-mix, a spectrogram-level variant, underperforms unless the original label is preserved. Lungmix builds a mask $M_{\text{Lungmix}} = M_{\text{loudness}} \oplus R$, where $M_{\text{loudness}} = \lambda \cdot (M_i \oplus M_j)$ with $M_i$ marking the loud outliers of waveform $i$ and $R$ a random binary mask, so rare crackle and wheeze transients stay intact while the rest of the signal is plausibly mixed. The labels are one-hot vectors over the abnormal classes combined by bitwise OR, which turns multi-label mixtures into a label-powerset class, and the loss is cross-entropy against that OR label plus a rescaled Mixup term. With an Audio Spectrogram Transformer, the non-linear variant of Lungmix gives the best combined-test score for the ICBHI source (63.30) and the HF source (72.11, versus 68.56 without Mixup), and the paper reports that the HF-trained model approaches the score of a model trained directly on ICBHI.

Load-bearing premise

The headline gain is measured by the ICBHI average score, which pools crackle, wheeze, and both into a single abnormal group, and the best variant is chosen on the combined test set; if the goal is fine-grained abnormal-sound discrimination on unseen data, the tables do not show that.

Editorial extensions

If this is right

  • With HF as the source domain, Lungmix scores 72.11 on the combined test set versus 68.56 without Mixup, the paper's largest reported gain.
  • Dropping the loudness mask removes the improvement, so the loudness mask is the operative component, not random mixing alone.
  • For the SPR source, stacking Lungmix with Patch-mix produces the best combined-test score (64.07), indicating the two augmentations are compatible.
  • Patch-mix with label preservation outperforms Patch-mix with the Mixup loss, supporting the paper's claim that linear label interpolation mismatches sparse abnormal sounds.
  • Across all three source domains, Lungmix variants hold or improve most combined-test scores, whereas each prior Mixup baseline hurts at least one target domain.

Reading between the lines

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

  • The loudness-mask-plus-OR-label recipe is a natural fit for any audio task where the target events are sparse and loud against a quiet background, such as cough detection or animal call classification, but the paper does not test those settings.
  • Because the ICBHI average score pools crackle, wheeze, and both into one abnormal group, the reported gains may mostly reflect better normal-versus-abnormal separation; whether Lungmix also sharpens discrimination among abnormal sound types is not established by the paper's tables.
  • The bitwise-OR label scheme creates $2^{n-1}+1$ classes once $n$ abnormal types are considered; applying Lungmix to datasets with many annotation categories would likely require per-class binary heads or a hierarchical label structure.
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

3 major / 6 minor

Summary. The paper proposes Lungmix, a Mixup-style data augmentation method for respiratory sound classification under single-source domain generalization. Lungmix mixes two waveforms with a mask formed from a loudness-based mask and a random mask, and interpolates labels through a bitwise OR (or label preservation) rather than linear interpolation. The method is evaluated with an Audio Spectrogram Transformer on three public datasets (ICBHI, SPR, HF) and compared against several Mixup variants on held-out test domains. The authors report that Lungmix improves the ICBHI-style average score by up to 3.55 percentage points over the no-mixup baseline (HF source, COMB test, Lungmix non-linear variant) and achieves performance close to models trained directly on the target dataset.

Significance. If the empirical claims held, Lungmix would be a simple and useful contribution: it is a waveform-level augmentation that requires no auxiliary model, it targets an important practical problem (cross-dataset generalization in respiratory sound analysis), and it is evaluated across three public datasets with several baselines. The paper also advances a reasonable hypothesis that sparse abnormal events in respiratory sounds need structured mixing rather than global convex interpolation. However, the current evidence is not yet convincing: the best variant is selected on the test set used for the headline comparison, no repeated runs or error bars are reported, and the evaluation metric does not actually measure four-class discrimination. With additional experiments using a validation split and per-class metrics, the method could be a solid contribution; as written, the central generalization claim is under-supported.

major comments (3)
  1. [Section III-C, Table I] The central claim that Lungmix significantly enhances generalization is not supported by the reported evaluation protocol. The paper states explicitly in Section III-C that for the ICBHI source it "chose the best result on the COMB dataset," and the bolded results in Table I indicate that the best variant is selected on the combined test set for each source domain. Since several Lungmix variants are evaluated and the best one is then reported on the same test set, the headline improvement (HF source, COMB test: 72.11 vs 68.56, +3.55) is a selected maximum rather than the performance of a fixed method. No validation split or repeated-seed variance is reported, so the gap could be within run-to-run noise for a fine-tuned transformer on datasets of this size. Moreover, Table I contains held-out transfers where Lungmix is worse than the no-mixup baseline (e.g., source SPR, test HF: 53.93 vs 59.29), which further weakens the claim of uniform improvement. I recommend reporting results for a fixed variant chosen on a validation set and giving means and standard deviations over multiple seeds.
  2. [Eq. (11), Table II] The reported "4-class score" is not a four-class metric. Equation (11) defines sensitivity as (CC + CW + CB)/(NC + NW + NB), pooling crackle, wheeze, and both into a single abnormal group, and specificity as CN/NN for the normal class only. Consequently, the average score can improve without any improvement in four-class discrimination, for example by shifting errors between abnormal classes or by trading sensitivity for specificity. Table II shows precisely such a tradeoff: for the ICBHI source, the non-linear variant compared to the linear variant drops sensitivity from 61.67 to 47.51 while raising specificity from 60.21 to 79.08; the average score improves even though abnormal-sound detection worsens. To support the claim of improved four-class classification, please report per-class sensitivity and confusion matrices, or use a genuine multiclass metric such as macro-F1 or class-balanced accuracy.
  3. [Eqs. (3)--(8)] The mask construction is under-specified and difficult to reproduce precisely. Equation (5) defines Mi as a boolean mask, Eq. (6) multiplies it by λ to produce MLoudness with values in {0, λ}, Eq. (7) produces a binary random mask R, and Eq. (8) combines MLoudness and R with ⊕ (OR) without defining how OR acts on values of different scales; the text also claims MLungmix ∈ {0, λ, 1}, but the result when MLoudness = λ and R = 1 is not defined. In addition, Eq. (4) uses ⊕ for label interpolation and Eq. (8) uses it for mask combination, overloading the same symbol. Please provide explicit element-wise definitions (e.g., threshold-based binarization followed by logical OR, or an arithmetic max operation) for the final mask.
minor comments (6)
  1. [Section III-B] "Traing Settings" should be "Training Settings."
  2. [Figure 2 caption] The caption "An a crackle" is grammatically incorrect; it should be "A crackle and a wheeze are mixed into both."
  3. [Section III-A] The abbreviation COMB is used without prior definition; please define the combined test set in the dataset description or at first use.
  4. [Eq. (10)] The sentence "λ2 is rescaled LMixup according to the first loss" is vague; please specify the exact scaling rule for λ2.
  5. [Reference [17]] Reference [17] lists only the first author of Mixup; the full author list should be provided.
  6. [Section III-B] "origin implementation" should be "original implementation."

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: Lungmix is an empirical augmentation study whose claims rest on held-out benchmark comparisons, not on a derivation that reduces to its own inputs.

full rationale

The paper makes no analytic derivation that could be circular; its central claims are empirical comparisons in Table I (Lungmix non-linear vs. w/o Mixup) on unseen test domains. The Lungmix construction (Eqs. 3-8) and label interpolation (Eqs. 9-10) are stipulated modeling choices, not outputs derived from the results, and the generalization claim is checked against external datasets (ICBHI, SPR, HF) rather than being equivalent to the construction. The potential concern that the headline 3.55% gain is the best of nine variants selected on the combined test set is a test-set selection / soundness risk, not circularity, because the selected configuration is still evaluated on data it was not trained on and the comparison is an empirical measurement. The score in Eq. (11) pooling crackle, wheeze, and both into one sensitivity is likewise a metric-validity concern (Table II shows Se/Sp trade-offs), not a self-referential reduction: the score does not define the labels or the augmentation. No load-bearing self-citation or uniqueness theorem is invoked; references to Mixup, AST, and dataset papers are independent support. The paper even reports a non-improvement for Lungmix (w/o loudness) and notes that non-linear interpolation does not significantly enhance Se or Sp in the ablation, which cuts against any claim that the method's construction guarantees its reported score. Score 0 reflects absence of circularity; correctness and selection concerns are separate.

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

The method itself is a constructed augmentation; the central claim rests on several hand-chosen constants (λ, threshold, random probability, loss weights) and domain assumptions about what loudness and OR labels mean. No new physical entities are introduced.

free parameters (4)
  • Beta distribution shape α (λ ~ Beta(α, α)) = not reported
    Controls how often the mask is near 0 or 1; material to the mixing behavior but never specified in the paper.
  • Loudness threshold multiplier 2·std (Eq. 5) = 2
    The mask relies on mean + 2·std, a hand-chosen constant that defines what counts as a meaningful or loud region; no sensitivity analysis.
  • Random mask probability 0.5 (Eq. 7) = 0.5
    Each sample position is assigned to one waveform with probability 0.5; the paper does not vary or justify this value.
  • Loss weights λ1, λ2 (Eq. 10) = λ1=1, λ2 rescaled (value not specified)
    The combination of OR-label CE and standard mixup loss is controlled by λ2, whose re-scaling rule is described only verbally.
assumptions (4)
  • ad hoc to paper The loudness mask M_i = |x_i| > |mean(x_i) + 2·std(x_i)| separates diagnostically meaningful waveform events from irrelevant parts.
    Invoked in Eq. (5) and used to construct MLoudness; if loud regions do not correspond to crackles or wheezes, the augmentation introduces label noise.
  • domain assumption Any mixture containing an abnormal component is labeled abnormal (OR semantics), e.g., normal+crackle -> crackle.
    Stated in Section II-C; this semantic rule is central to label interpolation and may over-credit faint abnormal components.
  • domain assumption The aggregate score in Eq. (11), averaging recall over all abnormal classes and specificity over normal, is a faithful measure of four-class classification quality.
    The metric collapses crackle, wheeze, and both into one abnormal pool and ignores inter-abnormal confusion; the reported improvements are in this collapsed score.
  • domain assumption AST fine-tuned on source data with the shown hyperparameters is a valid proxy for general respiratory sound classifiers.
    The entire evaluation uses one architecture family (AST) with one pretraining; generalization of the augmentation to other backbones is untested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lungmix: A Mixup-Based Strategy for Generalization in Respiratory Sound Classification." pith.science (2026). https://pith.science/paper/Q67KU3MQ

@misc{pith2026250100064,
  author       = {Pith},
  title        = {Pith review of: Lungmix: A Mixup-Based Strategy for Generalization in Respiratory Sound Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q67KU3MQ}},
  note         = {Machine review of arXiv:2501.00064}
}
read the original abstract

Respiratory sound classification plays a pivotal role in diagnosing respiratory diseases. While deep learning models have shown success with various respiratory sound datasets, our experiments indicate that models trained on one dataset often fail to generalize effectively to others, mainly due to data collection and annotation \emph{inconsistencies}. To address this limitation, we introduce \emph{Lungmix}, a novel data augmentation technique inspired by Mixup. Lungmix generates augmented data by blending waveforms using loudness and random masks while interpolating labels based on their semantic meaning, helping the model learn more generalized representations. Comprehensive evaluations across three datasets, namely ICBHI, SPR, and HF, demonstrate that Lungmix significantly enhances model generalization to unseen data. In particular, Lungmix boosts the 4-class classification score by up to 3.55\%, achieving performance comparable to models trained directly on the target dataset.

Figures

Figures reproduced from arXiv: 2501.00064 by the authors.

Figure 1
Figure 1. Performance comparison of simple fine-tuned audio spectrogram [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Visualization of Lungmix. An a crackle and wheeze are mixed into both. The grey parts denote the random mask, and the white parts denote the loudness mask. The zoomed-in section highlights the short and discontinuous crackle sound. The part under the zoomed-in is randomly generated padding. It combines two waveforms xi and xj using a mask MLungmix ∈ {0, λ, 1} 1×maxlen(xi,xj ) comes from the random mask and the loudn… view at source ↗
Figure 3
Figure 3. Visualization of label interpolation. (1) is linear interpolation, (2) is [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

  1. [1]

    Fundamentals of lung auscultation,

    A. Bohadana, G. Izbicki, and S. S. Kraman, “Fundamentals of lung auscultation,” New England Journal of Medicine , vol. 370, no. 8, pp. 744–751, 2014. [Online]. Available: https://www.nejm.org/doi/full/10. 1056/NEJMra1302901

    work page 2014

  2. [2]

    An open access database for the evaluation of respiratory sound classification algorithms,

    B. M. Rocha, D. Filos, L. Mendes, G. Serbes, S. Ulukaya, Y . P. Kahya, N. Jakovljevic, T. L. Turukalo, I. M. V ogiatzis, E. Perantoni, E. Kaimakamis, P. Natsiavas, A. Oliveira, C. J ´acome, A. Marques, N. Maglaveras, R. Pedro Paiva, I. Chouvarda, and P. de Carvalho, “An open access database for the evaluation of respiratory sound classification algorithms...

    work page 2019

  3. [3]

    Sprsound: Open-source sjtu paediatric respiratory sound database,

    Q. Zhang, J. Zhang, J. Yuan, H. Huang, Y . Zhang, B. Zhang, G. Lv, S. Lin, N. Wang, X. Liu, M. Tang, Y . Wang, H. Ma, L. Liu, S. Yuan, H. Zhou, J. Zhao, Y . Li, Y . Yin, L. Zhao, G. Wang, and Y . Lian, “Sprsound: Open-source sjtu paediatric respiratory sound database,” IEEE Transactions on Biomedical Circuits and Systems , vol. 16, no. 5, pp. 867–881, 2022

    work page 2022

  4. [4]

    Bench- marking of eight recurrent neural network variants for breath phase and adventitious sound detection on a self-developed open-access lung sound database—hf lung v1,

    F.-S. Hsu, S.-R. Huang, C.-W. Huang, C.-J. Huang, Y .-R. Cheng, C.- C. Chen, J. Hsiao, C.-W. Chen, L.-C. Chen, Y .-C. Lai et al. , “Bench- marking of eight recurrent neural network variants for breath phase and adventitious sound detection on a self-developed open-access lung sound database—hf lung v1,” PLoS One , vol. 16, no. 7, p. e0254134, 2021

    work page 2021

  5. [5]

    Machine learning in lung sound analysis: A systematic review,

    R. Palaniappan, K. Sundaraj, and N. U. Ahamed, “Machine learning in lung sound analysis: A systematic review,” Biocybernetics and Biomedical Engineering , vol. 33, no. 3, pp. 129–135, 2013. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0208521613000168

    work page 2013

  6. [6]

    Lungrn+nl: An improved adventitious lung sound classification using non-local block resnet neural network with mixup data augmentation,

    Y . Ma, X. Xu, and Y . Li, “Lungrn+nl: An improved adventitious lung sound classification using non-local block resnet neural network with mixup data augmentation,” in Interspeech, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:226206978

    work page 2020

  7. [7]

    RespireNet: A Deep Neural Network for Accurately Detecting Abnormal Lung Sounds in Limited Data Setting

    S. Gairola, F. Tom, N. Kwatra, and M. Jain, “Respirenet: A deep neural network for accurately detecting abnormal lung sounds in limited data setting,” CoRR, vol. abs/2011.00196, 2020. [Online]. Available: https://arxiv.org/abs/2011.00196

    work page Pith review arXiv 2011

  8. [8]

    Adventitious Respiratory Classification Using Attentive Residual Neu- ral Networks,

    Z. Yang, S. Liu, M. Song, E. Parada-Cabaleiro, and B. W. Schuller, “Adventitious Respiratory Classification Using Attentive Residual Neu- ral Networks,” in Proc. Interspeech 2020 , 2020, pp. 2912–2916

    work page 2020

Show all 31 references
  1. [9]

    Cnn-moe based framework for classification of respiratory anomalies and lung disease detection,

    L. Pham, H. Phan, R. Palaniappan, A. Mertins, and I. McLoughlin, “Cnn-moe based framework for classification of respiratory anomalies and lung disease detection,” IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 8, pp. 2938–2947, 2021

    2021

  2. [10]

    AST: audio spectrogram transformer,

    Y . Gong, Y . Chung, and J. R. Glass, “AST: audio spectrogram transformer,” CoRR, vol. abs/2104.01778, 2021. [Online]. Available: https://arxiv.org/abs/2104.01778

    2021 arXiv

  3. [11]

    Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification,

    S. Bae, J.-W. Kim, W.-Y . Cho, H. Baek, S. Son, B. Lee, C. Ha, K. Tae, S. Kim, and S.-Y . Yun, “Patch-mix contrastive learning with audio spectrogram transformer on respiratory sound classification,” in INTERSPEECH 2023 , ser. interspeech 2023. ISCA, Aug. 2023. [Online]. Avail...

    2023 doi

  4. [12]

    Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,

    J.-W. Kim, M. Toikkanen, S. Bae, M. Kim, and H.-Y . Jung, “Repaugment: Input-agnostic representation-level augmentation for respiratory sound classification,” 2024. [Online]. Available: https: //arxiv.org/abs/2405.02996

    2024 arXiv

  5. [13]

    Bts: Bridging text and sound modalities for metadata- aided respiratory sound classification,

    J.-W. Kim, M. Toikkanen, Y . Choi, S.-E. Moon, and H.-Y . Jung, “Bts: Bridging text and sound modalities for metadata- aided respiratory sound classification,” 2024. [Online]. Available: https://arxiv.org/abs/2406.06786

    2024 arXiv

  6. [14]

    Masked modeling duo: Towards a universal audio pre-training framework,

    D. Niizumi, D. Takeuchi, Y . Ohishi, N. Harada, and K. Kashino, “Masked modeling duo: Towards a universal audio pre-training framework,” 2024. [Online]. Available: https://arxiv.org/abs/2404.06095

    2024 arXiv

  7. [15]

    Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound classification,

    J.-W. Kim, S. Bae, W.-Y . Cho, B. Lee, and H.-Y . Jung, “Stethoscope- guided supervised contrastive learning for cross-domain adaptation on respiratory sound classification,” 2023. [Online]. Available: https://arxiv.org/abs/2312.09603

    2023 arXiv

  8. [16]

    A domain transfer based data augmentation method for automated respiratory classification,

    Z. Wang and Z. Wang, “A domain transfer based data augmentation method for automated respiratory classification,” in ICASSP 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2022, pp. 9017–9021

    2022

  9. [17]

    mixup: Beyond empirical risk minimization,

    H. Zhang, “mixup: Beyond empirical risk minimization,” arXiv preprint arXiv:1710.09412, 2017

    2017 arXiv

  10. [18]

    How does mixup help with robustness and generalization?

    L. Zhang, Z. Deng, K. Kawaguchi, A. Ghorbani, and J. Zou, “How does mixup help with robustness and generalization?” 2021. [Online]. Available: https://arxiv.org/abs/2010.04819

    2021 arXiv

  11. [19]

    Gap-aug: Gamma patch-wise correction augmentation method for respiratory sound classification,

    A.-Y . Chang, J.-T. Tzeng, H.-Y . Chen, C.-W. Sung, C.-H. Huang, E. P.- C. Huang, and C.-C. Lee, “Gap-aug: Gamma patch-wise correction augmentation method for respiratory sound classification,” in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal...

    2024

  12. [20]

    Puzzle mix: Exploiting saliency and local statistics for optimal mixup,

    J.-H. Kim, W. Choo, and H. O. Song, “Puzzle mix: Exploiting saliency and local statistics for optimal mixup,” 2020. [Online]. Available: https://arxiv.org/abs/2009.06962

    2020 arXiv

  13. [21]

    Resizemix: Mixing data with preserved object information and true labels,

    J. Qin, J. Fang, Q. Zhang, W. Liu, X. Wang, and X. Wang, “Resizemix: Mixing data with preserved object information and true labels,” arXiv preprint arXiv:2012.11101, 2020

    2012 arXiv

  14. [22]

    Attention mixup: An accurate mixup scheme based on interpretable attention mechanism for multi-label audio classification,

    W. Liu, Y . Ren, and J. Wang, “Attention mixup: An accurate mixup scheme based on interpretable attention mechanism for multi-label audio classification,” in ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , 2023, pp. 1–5

    2023

  15. [23]

    Domain generalization: A survey,

    K. Zhou, Z. Liu, Y . Qiao, T. Xiang, and C. C. Loy, “Domain generalization: A survey,” IEEE Transactions on Pattern Analysis and Machine Intelligence , p. 1–20, 2022. [Online]. Available: http://dx.doi.org/10.1109/TPAMI.2022.3195549

    2022

  16. [24]

    Mining multi-label data,

    G. Tsoumakas, I. M. Katakis, and I. P. Vlahavas, “Mining multi-label data,” in Data Mining and Knowledge Discovery Handbook , 2010. [Online]. Available: https://api.semanticscholar.org/CorpusID:22998

    2010

  17. [25]

    Theoretical analysis of self-training with deep networks on unlabeled data,

    C. Wei, K. Shen, Y . Chen, and T. Ma, “Theoretical analysis of self-training with deep networks on unlabeled data,” ArXiv, vol. abs/2010.03622, 2020. [Online]. Available: https://api.semanticscholar. org/CorpusID:222209080

    2010 arXiv

  18. [26]

    Audio set: An ontology and human-labeled dataset for audio events,

    J. F. Gemmeke, D. P. W. Ellis, D. Freedman, A. Jansen, W. Lawrence, R. C. Moore, M. Plakal, and M. Ritter, “Audio set: An ontology and human-labeled dataset for audio events,” in Proc. IEEE ICASSP 2017 , New Orleans, LA, 2017

    2017

  19. [27]

    Imagenet large scale visual recognition challenge,

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein et al., “Imagenet large scale visual recognition challenge,” International journal of computer vision, vol. 115, pp. 211–252, 2015

    2015

  20. [28]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift,

    S. Ioffe and C. Szegedy, “Batch normalization: Accelerating deep network training by reducing internal covariate shift,” CoRR, vol. abs/1502.03167, 2015. [Online]. Available: http://arxiv.org/abs/1502. 03167

    2015 arXiv

  21. [29]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” in International Conference on Learning Representations , 2019. [Online]. Available: https://openreview.net/forum?id=Bkg6RiCqY7

    2019

  22. [30]

    Sgdr: Stochastic gradient descent with warm restarts,

    ——, “Sgdr: Stochastic gradient descent with warm restarts,” 2017. [Online]. Available: https://arxiv.org/abs/1608.03983

    2017 arXiv

  23. [31]

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

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Regularization strategy to train strong classifiers with localizable features,” 2019. [Online]. Available: https://arxiv.org/abs/1905.04899

    2019 arXiv

Pith tools

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