pith. sign in

arxiv: 2607.02254 · v1 · pith:YGH5B2VInew · submitted 2026-07-02 · 📡 eess.AS

Cross Domain Few-Shot Class-Incremental Audio Classification Via Adversarial Contrastive Learning

Yongjie Si , Yanxiong Li , Sen Huang , Beibei Liu This is my paper

Pith reviewed 2026-07-03 04:53 UTC · model grok-4.3

classification 📡 eess.AS
keywords few-shot class-incremental learningcross-domain audio classificationadversarial contrastive learningdomain shiftaudio classificationincremental learning
0
0 comments X

The pith

Adversarial contrastive training lets a frozen encoder and updated classifier classify new audio classes from shifted domains.

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

Standard few-shot class-incremental audio classification assumes base and incremental classes come from the same domain, yet domain shifts commonly occur between them. The paper proposes adversarial contrastive training to solve the resulting cross-domain problem by training an encoder only on base classes, freezing it thereafter, and training the classifier across all sessions. This setup allows the model to handle samples from unseen domains without explicit domain adaptation modules. Experiments across six pairs of cross-domain datasets report higher average accuracy than existing methods. Readers would care because the result makes incremental audio learning viable when recording conditions or devices change between training phases.

Core claim

The paper claims that a strategy of adversarial contrastive training enables the model to effectively classify samples of different classes from unseen domains in cross-domain few-shot class-incremental audio classification, where the encoder is trained in the base session but frozen in incremental sessions and the classifier is trained in all sessions, exceeding state-of-the-art methods in average accuracy on six pairs of cross-domain datasets.

What carries the argument

Adversarial contrastive training applied to an encoder-classifier model that freezes the encoder after base-session training.

If this is right

  • The classifier can adapt to new classes from different domains while the encoder remains fixed after base training.
  • Average accuracy improves over prior methods on multiple source-to-target domain pairs without extra adaptation components.
  • The approach applies directly to incremental audio tasks where base and new samples follow different distributions.
  • Training the classifier in every session while freezing the encoder reduces the risk of overwriting earlier knowledge.

Where Pith is reading between the lines

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

  • The method may lower deployment cost by avoiding full retraining when new domains appear.
  • Success across six dataset pairs suggests the training objectives produce features that transfer across common audio domain shifts.
  • The frozen-encoder design could be tested on incremental tasks in other sensor modalities that exhibit similar distribution changes.

Load-bearing premise

The domain shift between base and incremental class samples can be adequately addressed by adversarial contrastive training with a frozen encoder after the base session and an updated classifier, without needing explicit domain adaptation modules.

What would settle it

If the method recorded lower average accuracy than state-of-the-art approaches across the six pairs of cross-domain datasets, the central performance claim would be falsified.

Figures

Figures reproduced from arXiv: 2607.02254 by Beibei Liu, Sen Huang, Yanxiong Li, Yongjie Si.

Figure 1
Figure 1. Figure 1: Framework of our method. In base training, we learn domain-invariant embeddings via adversarial contrastive training using a spectral disruptor. In incremental training, we freeze the encoder and update the classifier with embeddings of new classes samples and mean vectors of old classes’ embeddings. 𝛼, 𝛽, 𝜆: three adjustable coefficients; ℓ௖௘ ௡௘௪: cross-entropy loss ℓ௖௘ on embeddings of new classes; ℓୡୣ ୭… view at source ↗
Figure 2
Figure 2. Figure 2: AA scores (in %) obtained by our method on the NS→LS using various values of N-way K-shot. In the second extended experiment, we use t-SNE [29] to show the effectiveness of the adversarial contrastive training strategy [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Distributions of testing embeddings of five source and target classes on the NS→LS (a) without using and (b) using the proposed adversarial contrastive training strategy. 4. Conclusions We solved the CD-FCAC problem via the adversarial training and supervised contrastive loss. Based on the description of our method and results, we can draw two conclusions. First, our method exceeds state-of-the-art methods… view at source ↗
read the original abstract

Current Few-shot Class-incremental Audio Classification (FCAC) methods assume that samples of base and incremental classes are in the same domain (following the same distribution). However, there is generally a domain shift between the above two types of samples. In this paper, we explore the problem of Cross Domain FCAC where samples of base and incremental classes have domain shift. We propose a strategy of adversarial contrastive training which enables the model to effectively classify samples of different classes from unseen domains. The model consists of an encoder and a classifier. The encoder is trained in base session but frozen in incremental sessions, whereas the classifier is trained in all sessions. Experiments are done on six pairs of cross-domain datasets. Results show that our method exceeds state-of-the-art methods in average accuracy. The code is at https://github.com/YongjieSi/ACL.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 2 minor

Summary. The paper introduces the Cross Domain Few-Shot Class-Incremental Audio Classification (FCAC) problem, where base and incremental classes exhibit domain shift. It proposes an adversarial contrastive learning strategy consisting of an encoder trained only on base-session data (then frozen) and a classifier updated across all sessions. The method is evaluated on six pairs of cross-domain audio datasets and reported to exceed state-of-the-art average accuracy; code is released at https://github.com/YongjieSi/ACL.

Significance. If the empirical superiority holds under detailed scrutiny, the work would address a practically relevant gap in audio classification by relaxing the same-domain assumption common in FCAC. Releasing code supports reproducibility and is a positive contribution.

major comments (2)
  1. [Method] Method section (architecture description): the encoder is trained solely on base data and frozen thereafter, with only the classifier updated in incremental sessions. No explicit domain-adversarial loss, feature alignment term, or encoder fine-tuning is described to mitigate distribution mismatch; the contrastive component therefore operates on potentially misaligned frozen representations. This assumption is load-bearing for the central claim that the strategy handles cross-domain shift.
  2. [Experiments] Experiments section: the abstract and results claim outperformance on six dataset pairs, yet no specific baselines, metrics (e.g., per-session accuracy with standard deviation), number of runs, statistical significance tests, or quantification of domain shift (e.g., via MMD or classifier accuracy on domain labels) are referenced. Without these, it is impossible to verify whether the reported average accuracy improvement is robust or merely reflects weak baselines.
minor comments (2)
  1. [Method] Notation for the adversarial contrastive loss should be formalized with an equation; the current prose description leaves unclear whether an adversarial objective is present or whether the term is used descriptively.
  2. [Experiments] The six dataset pairs should be explicitly listed with their domain characteristics (e.g., recording conditions, sampling rates) in a table to allow readers to assess the severity of the shifts.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will make revisions to improve clarity and completeness.

read point-by-point responses

  1. Referee: [Method] Method section (architecture description): the encoder is trained solely on base data and frozen thereafter, with only the classifier updated in incremental sessions. No explicit domain-adversarial loss, feature alignment term, or encoder fine-tuning is described to mitigate distribution mismatch; the contrastive component therefore operates on potentially misaligned frozen representations. This assumption is load-bearing for the central claim that the strategy handles cross-domain shift.

    Authors: The proposed adversarial contrastive training is the core mechanism intended to promote robustness to domain shift even with a frozen encoder, by structuring the contrastive objective during base-session training to support generalization to incremental domains. We acknowledge that the current description does not explicitly detail an additional domain-adversarial loss term. To address this, we will revise the method section to more clearly articulate how the adversarial contrastive strategy mitigates the distribution mismatch without requiring encoder updates or explicit alignment losses in incremental sessions. revision: yes

  2. Referee: [Experiments] Experiments section: the abstract and results claim outperformance on six dataset pairs, yet no specific baselines, metrics (e.g., per-session accuracy with standard deviation), number of runs, statistical significance tests, or quantification of domain shift (e.g., via MMD or classifier accuracy on domain labels) are referenced. Without these, it is impossible to verify whether the reported average accuracy improvement is robust or merely reflects weak baselines.

    Authors: We agree that the experimental reporting requires additional detail to substantiate the claims. In the revised manuscript, we will expand the experiments section to specify the baselines, report per-session accuracies with standard deviations over the number of runs performed, include statistical significance tests, and add quantification of domain shift (e.g., via MMD). revision: yes

Circularity Check

0 steps flagged

No circularity: empirical method proposal without derivations or self-referential predictions

full rationale

The paper presents an empirical ML method for cross-domain few-shot class-incremental audio classification using adversarial contrastive training. The architecture (encoder trained then frozen after base session, classifier updated across sessions) is described and evaluated directly on six pairs of cross-domain datasets, with results compared to SOTA. No equations, first-principles derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text. The central claim rests on experimental accuracy improvements rather than any chain that reduces to its own inputs by construction, rendering the work self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only abstract available; no details on free parameters, axioms, or invented entities can be extracted.

pith-pipeline@v0.9.1-grok · 5684 in / 1117 out tokens · 31951 ms · 2026-07-03T04:53:15.624182+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    It has widespread app lications, such as wildlife protection [1], road surveillance [2], classification of acoust ic scene [3], video analysis [4] and speaker analysis [5]

    Introduction Audio classification (AC) is a task to identify different audio classes. It has widespread app lications, such as wildlife protection [1], road surveillance [2], classification of acoust ic scene [3], video analysis [4] and speaker analysis [5]. Many AC methods [6]-[9] require abundant samples of predefined classes for model training. The tra...

    work page

  2. [2]

    We explore a new problem of CD-FCAC where the domain shift and class increment coexist

    work page

  3. [3]

    We propose an adversarial contrastive training strategy which is a fusion of adversarial training and contrastive learning

    work page

  4. [4]

    We propose a CD-FCAC me thod which exceeds state- of-the-art methods in average accuracy

    work page

  5. [5]

    Problem Definition The CD-FCAC consists of M sessions, including one base session (session 0) and ( M-1) incremental sessions (sessions 1 to (M-1))

    Method 2.1. Problem Definition The CD-FCAC consists of M sessions, including one base session (session 0) and ( M-1) incremental sessions (sessions 1 to (M-1)). we consider a user-friendly setting where the samples * Corresponding author: Yanxiong Li (eeyxli@scut.edu.cn). in base session are from a single source domain, while the samples in incremental se...

    work page

  6. [6]

    Select audio samples from 𝑫଴ ௧௥and extract its log-Mel spectrogram (𝑿௦,𝒀 ௦)

    work page

  7. [7]

    𝑿௧←Spectral disturbance on 𝑿௦

    While maximizing training do 4 . 𝑿௧←Spectral disturbance on 𝑿௦

    work page

  8. [8]

    𝑿௧←𝑿 ௧+ ζ𝑚𝑎𝑥∇𝑿೟(ℓ௖௘((𝑿௧,𝒀 ௦); 𝜃) − 𝛾𝑑((𝑿௧,𝒀 ௦), (𝑿௦,𝒀 ௦)))

    F o r R epochs do 6 . 𝑿௧←𝑿 ௧+ ζ𝑚𝑎𝑥∇𝑿೟(ℓ௖௘((𝑿௧,𝒀 ௦); 𝜃) − 𝛾𝑑((𝑿௧,𝒀 ௦), (𝑿௦,𝒀 ௦)))

    work page

  9. [9]

    𝑿= [𝑿௦,𝑿 ௧], 𝒀= [𝒀௦,𝒀 ௦]

    work page

  10. [10]

    𝜃 ← 𝜃 − ζ ௠௜௡∇ఏℓ௠௜௡(𝑿,𝒀 ;𝜃 )

    work page

  11. [11]

    Experimental Datasets Table 2: Detailed information of the LS-100/NSynth -100/FSC-89

    Experiments 3.1. Experimental Datasets Table 2: Detailed information of the LS-100/NSynth -100/FSC-89. Param D0 Dm (1 ≤ m ≤ (M-1)) 𝑫଴ ௧௥ 𝑫଴ ௧௘ 𝑫௠ ௧௥ 𝑫௠ ௧௘ #C 60/55/59 60/55/59 40/45/30 40/45/30 L(h) 16.66/12.23 /13.11 3.33/1.52 /3.28 11.11/5.00 /4.17 2.22/5.00 /1.67 #S/C 500/200/800 100/100/200 500/100/500 100/100/200 #C: number of classes, L(h): Length(h...

    work page

  12. [12]

    Based on the description of our method and results, we can draw two conclusions

    Conclusions We solved the CD-FCAC problem via the adversarial training and supervised contrastive loss. Based on the description of our method and results, we can draw two conclusions. First, our method exceeds state-of-the-art methods in average accuracy. Second, the proposed adversarial contrastive training strategy benefits to improve the model’s perfo...

    work page

  13. [13]

    Acknowledgements This work was supported by the national natural science foundation of China (62371195, 62111530145, 61771200), the exchange project of the 10th Meeting of the China-Croatia Science and Technology Cooperation Committee (No. 10-34)

    work page

  14. [14]

    The authors take full responsibility for the content

    Generative AI Use Disclosure No generative AI tools (e.g., LLM/ChatGPT) were used to produce any part of this work, except that a limited amount of grammar/spelling checking may have been performed on the final text using standard to ols. The authors take full responsibility for the content

    work page

  15. [15]

    Comparison of feature extraction methods for sound-based classification of honey bee activity,

    A. Terni, N. Ortolani, I. Nolasco, E. Bonitos, and S. Cecchi, “Comparison of feature extraction methods for sound-based classification of honey bee activity,” IEEE/ACM TASLP, vol. 30, pp. 112-122, 2022

    work page 2022

  16. [16]

    Anomalous sou nd detection using deep audio repr esentation and a BLSTM network for audio surveillance of roads,

    Y. Li, X. Li, Y. Zhang, M. Liu, and W. Wang, “Anomalous sou nd detection using deep audio repr esentation and a BLSTM network for audio surveillance of roads,” IEEE Access, vol. 6, pp. 58043- 58055, 2018

    work page 2018

  17. [17]

    Low-compl exity acoustic scene classification us ing parallel attention-convolut ion network,

    Y. Li, J. Tan, G. Chen, J. Li, Y. Si, and Q. He, “Low-compl exity acoustic scene classification us ing parallel attention-convolut ion network,” in Proc. of Interspeech, 2024, pp. 1-5

    work page 2024

  18. [18]

    Detecting video anomalies by jointly util izing appearance and skeleton information,

    W. Pang, et al., “Detecting video anomalies by jointly util izing appearance and skeleton information,” Expert Syst. Appl., vol. 246, pp. 1-12, 2024

    work page 2024

  19. [19]

    Speaker clust ering by co-optimizing deep representation learning and cluster estimation,

    Y. Li, W. Wang, M. Liu, Z. Jiang, and Q. He, “Speaker clust ering by co-optimizing deep representation learning and cluster estimation,” IEEE TMM, vol. 23, pp. 3377-3387, 2021

    work page 2021

  20. [20]

    Learnable counterfactual attention for music classification,

    Y.-X. Lin, J.-C. Lin, W.-L. Wei and J.-C. Wang, “Learnable counterfactual attention for music classification,” IEEE TASLP, vol. 33, pp. 570-585, 2025

    work page 2025

  21. [21]

    Classification of urban sound using sequential convolutional neural network model and its visualisation,

    M. Agarwal, K. S. Gill, S. Chattopadhyay and M. Singh, “Classification of urban sound using sequential convolutional neural network model and its visualisation,” in Proc. of ICITEICSI, 2024, pp. 1-5. [ 8 ] K . M . L i m , C . P . L e e , Z . Y . L e e , J . Y . L i m a n d J . N . M o g a n , “Acoustic event classification with enhanced efficientNet,” in ...

    work page 2024

  22. [22]

    Enhancing conformer-based sound event detection using frequency dynamic convolutions a nd beats audio embeddings,

    S. Barahona, D. de Benito-Gorrón, D. T. Toledano and D. Ram os, “Enhancing conformer-based sound event detection using frequency dynamic convolutions a nd beats audio embeddings,” IEEE/ACM TASLP, vol. 32, pp. 3896-3907, 2024

    work page 2024

  23. [23]

    Few shot continual learning for audio classification,

    Y. Wang, N. J. Bryan, M. Cartwright, J. P. Bello, and J. Salomon, “Few shot continual learning for audio classification,” in Proc. of IEEE ICASSP, 2021, pp. 321-325

    work page 2021

  24. [24]

    Few-shot class- incremental audio classifica tion using dynamically expanded classifier with self-attention modified prototypes,

    Y. Li, W. Cao, W. Xian, J. Li, and E. Benito’s, “Few-shot class- incremental audio classifica tion using dynamically expanded classifier with self-attention modified prototypes,” IEEE TMM , vol.26, pp.1346-1360, 2024

    work page 2024

  25. [25]

    Few-shot class-incremental audio classification via discriminative prototype learning,

    W. Xie, et al., “Few-shot class-incremental audio classification via discriminative prototype learning,” Expert Syst. Appl. , vol. 225, 2023, Art. no. 120044

    work page 2023

  26. [26]

    Few-shot c lass-incremental audio classifi cation using adaptively-refined prototypes,

    W. Xie, et al., “Few-shot c lass-incremental audio classifi cation using adaptively-refined prototypes,” in Proc. of Interspeech, 2023, pp. 301-305. [ 1 4 ] Y . L i , W . C a o , J . L i , W . X i e , a n d Q . H e , “ F e w - s h o t c l a s s - incremental audio classification using stochastic classifier,” in Proc. of Interspeech, 2023, pp. 4174-4178

    work page 2023

  27. [27]

    Few-shot class-incremental audio classification with adap tive mitigation of forgetting and overfitting,

    Y. Li, J. Li, Y. Si, J. Tan, and Q. He, “Few-shot class-incremental audio classification with adap tive mitigation of forgetting and overfitting,” IEEE/ACM TASLP, vol. 32, pp. 2297-2311, 2024

    work page 2024

  28. [28]

    Fully few-shot class-incremental audio classifi cation with adaptive improvemen t of stability and plasticity,

    Y. Si, Y. Li, J. Tan, G. Chen, Q. Li, and M. Russo, “Fully few-shot class-incremental audio classifi cation with adaptive improvemen t of stability and plasticity,” IEEE TASLP, vol. 33, pp. 418-433, 2025

    work page 2025

  29. [29]

    Fully few-shot class- incremental audio classifica tion using multi-level embedding extractor and ridge regression classifier,

    Y. Si, Y. Li, J. Tan, Q. He, and I. Kwak, “Fully few-shot class- incremental audio classifica tion using multi-level embedding extractor and ridge regression classifier,” in Proc. of Interspeech , 2025, pp. 1318-1322

    work page 2025

  30. [30]

    Supervised contrastive learning,

    P. Khosla, P. Teterwak, C. Wang, A. Sarna, Y. Tian, P. Iso la, A. Maschinot, C. Liu, and D. Krishnan, “Supervised contrastive learning,” in Proc. of NIPS, 2020. pp. 18661-18673

    work page 2020

  31. [31]

    Certifying some distributional robustness with p rincipled adversarial training,

    A. Sinha, H. Namkoong, and J. Duchi, “Certifying some distributional robustness with p rincipled adversarial training, ” in Proc. of International Conference on Learning Representations , 2018, pp. 1-34

    work page 2018

  32. [32]

    Generalizing to unseen domains via adversarial da ta augmentation,

    R. Volpi, H. Namkoong, O. Sener, and J.C. Duchi, V. Murino, and S. Savarese, “Generalizing to unseen domains via adversarial da ta augmentation,” in Proc. of Advances in Neural Information Processing Systems, 2018, pp. 1-11

    work page 2018

  33. [33]

    Adversaria l feature augmentation for cros s- domain few-shot classification,

    Y. Hu, A.J. Ma, “Adversaria l feature augmentation for cros s- domain few-shot classification,” in Proc. of ECCV, 2022, pp. 20-37

    work page 2022

  34. [34]

    An adversarial meta-training framewo rk for cross-domain few-shot learning,

    P. Tian, and S. Xie, “An adversarial meta-training framewo rk for cross-domain few-shot learning," IEEE Transactions on Multimedia, vol. 25, pp. 6881-6891, 2023

    work page 2023

  35. [35]

    Network randomizatio n: A simple technique for generalization in deep reinforcement learning,

    K. Lee, K. Lee, J. Shin, and H. Lee, “Network randomizatio n: A simple technique for generalization in deep reinforcement learning,” in Proc. of ICLR, 2020, pp. 1-22

    work page 2020

  36. [36]

    Understanding the difficulty of training deep feedforward neur al networks,

    X. Glorot, and Y. Bengio, “Understanding the difficulty of training deep feedforward neur al networks,” in Proc. of the thirteenth international conference on artific ial intelligence and statistics , 2010, pp. 249-256

    work page 2010

  37. [37]

    Cross-domain Few-shot Learnin g with Task-specific Adapters,

    W. Li, X. Liu and H. Bilen, "Cross-domain Few-shot Learnin g with Task-specific Adapters," in Proc. of CVPR , 2022, pp. 7151- 7160,

    work page 2022

  38. [38]

    Revisiti ng Prototypical Network for Cross Domain Few-Shot Learning,

    F. Zhou, P. Wang, L. Zhang, W. Wei and Y. Zhang, "Revisiti ng Prototypical Network for Cross Domain Few-Shot Learning," in Proc. Of CVPR, 2023, pp. 20061-20070

    work page 2023

  39. [39]

    F ew- shot incremental learning with continually evolved classifiers,

    C. Zhang, N. Song, G. Lin, Y. Zheng, P. Pan, and Y. Xu, “F ew- shot incremental learning with continually evolved classifiers, ” in Proc. of CVPR, 2021, pp. 12450 -1245

    work page 2021

  40. [40]

    Cross-domain few-shot incremental learning for point-cloud recognition,

    Y. Tan, and X. Xiang, “Cross-domain few-shot incremental learning for point-cloud recognition,” in Proc. of IEEE/CVF Winter Conference on Applications of Computer Vision , 2024, pp. 2296- 2305

    work page 2024

  41. [41]

    Visualizing data using T- SNE,

    L. V. D. Maaten, and G. Hinton, “Visualizing data using T- SNE,” J. Mach. Learn. Res., vol. 9, no. 96, pp. 2579-2605, 2008

    work page 2008

  42. [42]

    Deep Residual Learnin g for Image Recognition,

    K. He, X. Zhang, S. Ren and J. Sun, "Deep Residual Learnin g for Image Recognition," in Proc. of IEEE CVPR, 2016, 1022 pp. 770- 778

    work page 2016

  43. [43]

    Networ k randomization: A simple technique for generalization in deep reinforcement learning

    Kimin Lee, Kibok Lee, Jinw oo Shin, and Honglak Lee. Networ k randomization: A simple technique for generalization in deep reinforcement learning. in Proc. of IEEE ICLR, 2020

    work page 2020