Pith. sign in

REVIEW 5 major objections 5 minor 58 references

Fully Few-shot Class-incremental Audio Classification Using Multi-level Embedding Extractor and Ridge Regression Classifier

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

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

Pith's one-line read Frozen multi-level embeddings and ridge regression outperform prior methods in fully few-shot continual audio classification.

desk verdict A sensible incremental FFCAC system, but the state-of-the-art claim is not yet demonstrated because baseline numbers and splits are uncontrolled. read the letter →

arxiv 2506.18406 v1 pith:XQMSVTPD submitted 2025-06-23 eess.AS

Yongjie Si , Yanxiong Li , Jiaxin Tan , Qianhua He , Il-Youp Kwak This is my paper · ORCID
classification eess.AS
keywords fullyfew-shotclass-incrementallearningaudioclassificationSpectrogramTransformermulti-levelembeddingfusionridgeregressionclassifiercontinual
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

Fully few-shot class-incremental audio classification (FFCAC) asks a model to keep learning new sound classes when every class, including the base classes, has only a handful of training samples. The paper argues that this can be handled by decoupling the model into a multi-level embedding extractor and a ridge-regression classifier. The extractor, built on an Audio Spectrogram Transformer, is finetuned in the base session and then frozen; the classifier is updated in each incremental session by a closed-form ridge-regression solution. On LS-100, NSynth-100, and FSC-89, the method reports higher average accuracy and lower performance degradation than all compared prior methods, with the largest gain over the closest baseline on NSynth-100. If the result holds, it would show that a simple frozen-extractor recipe can replace more complex dynamically expanded classifiers for few-shot continual audio learning.

What carries the argument

The central objects are the Multi-level Embedding Extractor (MEE) and the Ridge Regression Classifier (RRC). MEE is an Audio Spectrogram Transformer whose block outputs are concatenated and reweighted by an MLP-plus-Softmax fusion module; this fuses local, concrete features from shallow blocks with global, abstract features from deep blocks into a single embedding. RRC computes classifier weights as $W=(E^T E+\lambda I)^{-1}E^T Y$, maintaining the Gram matrix $G=E^T E$ and $C=E^T Y$ across sessions and updating them additively, which prevents catastrophic forgetting without retraining and reduces correlations between embedding dimensions to counter overfitting under few samples.

What would settle it

Rerun the same 5-way, 5-shot, ten-session protocol on a fresh random selection of 50 classes from LS-100, NSynth-100, or FSC-89, with the same frozen-extractor and ridge-regression code; if average accuracy does not reproduce the reported margins over AISP (roughly +3.7, +7.3, and +0.3 points), the result depends on the particular class draw or evaluation protocol.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a feature extractor pretrained on large-scale audio, lightly finetuned on only 25 base samples, and then frozen for all incremental sessions is enough to make a fully few-shot continual audio classifier state of the art. The embeddings come from fusing the features of several transformer blocks, and the classifier is a ridge-regression head whose weights are solved analytically, so each new session only adds to two accumulated matrices rather than retraining the network. Reported average accuracy is 58.95% on LS-100, 68.65% on NSynth-100, and 33.27% on FSC-89, exceeding the closest prior method AISP by about 3.7, 7.3, and 0.3 points, with lower performance degradation in every case.

Load-bearing premise

The method's key assumption is that a pretrained AST encoder, finetuned on only 25 base samples and then frozen, keeps producing discriminative embeddings for all future incremental classes.

Editorial extensions

If this is right

  • New sound classes can be added by updating two small matrices, so each incremental session is cheap in computation and memory relative to attention-based expansion modules.
  • The method's performance degradation over ten sessions is lower than that of every compared method, so old-class recognition degrades less as new classes arrive.
  • Fusion of multi-block features is a measurable ingredient: on NSynth-100, adding the fusion module to the ridge-regression classifier raises average accuracy from 65.31% to 68.65% in the paper's ablation.
  • Because the extractor is frozen after the base session, the method's complexity sits below several prior fully few-shot and few-shot continual methods while exceeding them in accuracy.

Reading between the lines

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

  • Editorial inference: the recipe is not tied to AST specifically; if another pretrained audio encoder yields embeddings with similar or better transferability, the same ridge-regression update should carry the gain, which is directly testable by swapping the encoder.
  • Editorial inference: the success on only 25 base samples suggests the bottleneck in FFCAC is classifier calibration rather than feature learning, so pretrained encoders with stronger general-purpose audio representations could shift results further without architectural changes.
  • Editorial inference: the additive G and C updates assume class-disjoint sessions; a natural extension outside the paper is to test whether the same update remains stable when sessions overlap in label space or when classes reappear.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper addresses Fully Few-shot Class-incremental Audio Classification (FFCAC), where both base and incremental sessions provide only a few training samples per class. The proposed method decouples the model into a multi-level embedding extractor (MEE), consisting of an audio spectrogram Transformer encoder plus a trained fusion module, and a ridge regression classifier (RRC) updated by a closed-form least-squares solution. The MEE is finetuned in the base session and frozen in incremental sessions; the RRC accumulates Gram-matrix-like terms G and C and recomputes classifier weights after each session. Experiments on LS-100, NSynth-100, and FSC-89 are reported with average accuracy (AA) and performance degradation (PD) metrics. An ablation on NSynth-100 shows that both the fusion module and the RRC improve the results. The main claimed contribution is that the method exceeds nine prior methods in AA and PD on the three datasets while having middling complexity. The manuscript includes code and describes the training configuration in detail.

Significance. The paper's recipe — a frozen pretrained transformer with multi-level feature fusion and a non-parametric ridge-regression classifier — is simple, coherent, and, if the numbers hold, would be a strong baseline for FFCAC. The ablation study in Table 3 is informative and supports the contribution of each component, and the closed-form classifier update is a practical way to avoid repeated optimization in incremental sessions. The release of code is helpful. However, the empirical claim is currently not demonstrated: the comparison against prior methods is not shown to be protocol-controlled, no run-to-run variability is reported, and one session's accuracy in Table 4 is anomalous. These issues are load-bearing because the paper's headline is an empirical performance claim.

major comments (5)
  1. [§3.1 and Table 1] Table 1 lists the full datasets as 100/100/89 classes, while §3.1 states that fifty classes are randomly selected from each dataset; the paper never specifies the 50-class subset, the random seed, the class ordering, or the session assignment. Because the incremental order is known to affect few-shot continual learning results, the reported comparisons in Tables 4-6 cannot be reproduced or checked for protocol match, and this affects the paper's main claim.
  2. [§3.4 and Tables 4–6] The baseline numbers are quoted without any statement that they were rerun under this paper's exact 50-class split, incremental order, and evaluation protocol. The closest baseline, AISP [39], is the authors' own prior work, and the earlier FFCAC method [38] is omitted from the comparison tables entirely. Without a controlled reimplementation or a detailed protocol-match statement, the reported AA advantages (0.34–7.32 points) may reflect split or ordering differences rather than method superiority.
  3. [§3.2 and Tables 4–6] Section 3.2 says that the final result is the average of 100 tests, but the paper reports no standard deviation, confidence interval, or per-seed values. Given the random 50-class selection and random session order, the FSC-89 AA gap of 0.34 points between Ours and AISP (33.27 vs. 32.93) is within plausible run-to-run variation, so the claim that the method 'exceeds all prior methods' is not statistically supported.
  4. [Table 4] On LS-100, the proposed method's accuracy jumps from 46.93 in session 8 to 57.36 in session 9, while AISP changes only from 47.13 to 48.60; adding one 5-class session should not normally cause a 10.43-point jump in overall accuracy. This anomaly suggests an evaluation-set inconsistency or a typo, and because session 9 contributes about one percentage point to the reported AA, it needs to be explained or corrected.
  5. [§2.2 and §2.3] The AST encoder is initialized from a pretrained model and then frozen after the base session, but the paper does not state what data the pretrained model was trained on or whether its training classes overlap with the evaluation datasets. If the pretrained model has already seen the incremental classes, the 'fully few-shot' protocol is compromised; at minimum, the pretraining corpus and any class overlap should be reported before the few-shot claim can be interpreted.
minor comments (5)
  1. [§2.3] Section 2.3 contains typos: 'characterisitcs' and 'incoporate' should be corrected.
  2. [Eq. (1)] Equation (1) uses W_{y_k} before defining it; the text should clarify that W_{y_k} is the classifier weight vector for the class of the kth sample.
  3. [§2.4] Section 2.4 says λ is obtained by cross-validation using training samples; the paper should state whether λ is tuned on the base session only or on all sessions and whether the same value is used throughout.
  4. [§2.4 and inference] The inference step uses cosine similarity between embeddings and classifier weights, but Eq. (5) computes weights with unnormalized embeddings; the paper should state whether embeddings and weights are normalized before the cosine operation.
  5. [Table 7] Table 7 reports different MACs and NP values for Ours across datasets, but the text does not explain that the input LMS sizes differ across datasets; a sentence noting this would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the claimed method gains rest on held-out empirical comparisons, not on fitting or self-citation.

full rationale

The paper's derivation chain is self-contained and non-circular. The classifier weights in Eq. (5)-(7) are a closed-form ridge regression solution using embeddings and one-hot labels, with matrices G and C updated by simple accumulation in Eqs. (8)-(9). The embeddings in Eqs. (3)-(4) come from the AST encoder and the fusion module, not from the classifier or from the reported accuracies. The only hyperparameter λ is obtained by cross-validation on training samples (Section 2.4) and does not manufacture held-out predictions; the reported AA and PD values in Section 3.2 are averages over test sets of current and prior sessions. The ablation study in Table 3 is a within-method comparison, and even there the conclusion is drawn from measured accuracies rather than from any equation that defines the output in terms of the claim. The principle self-citations are to the authors' prior FFCAC task definitions and AISP baseline ([38], [39]); these are disclosed prior work and the baseline numbers in Tables 4-6 are treated as external comparison values, not as a load-bearing proof step. The possible concern that prior baseline numbers may not share the exact 50-class split or evaluation protocol is an experimental validity issue, not circularity: it does not make a prediction reduce by construction to an input. Accordingly, no circularity step is identifiable, and the score is 0.

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

The method leans on a pretrained AST for transferable features, a plausible but untested domain assumption; lambda and eta are tuned or fixed values. The remaining assumptions are standard math or experimental comparability, and no invented entities are postulated.

free parameters (2)
  • Ridge regularization parameter lambda = obtained by cross-validation, value not reported
    In Eq. (5), lambda controls the closed-form classifier; it is tuned on training samples and affects all reported accuracies.
  • Cosine softmax scale factor eta = 16 (Table 2)
    Eq. (1) uses eta=16 in base-session finetuning; not derived, chosen by hand.
assumptions (4)
  • domain assumption AST pretrained on AudioSet transfers to target audio datasets
    Section 2.2 initializes the encoder with a pretrained AST; the method relies on these features remaining useful when only 25 base samples are available for finetuning.
  • domain assumption Features from different AST encoder blocks capture complementary levels of time-frequency abstraction
    Section 2.3 motivates fusing multi-level features; if blocks were redundant, the fusion module would add little.
  • standard math Closed-form ridge regression with accumulating G and C exactly updates the least-squares classifier for all seen samples
    Eqs. (5)-(9) assume the objective is a sum of squared errors over fixed embeddings, so incremental sums are sufficient.
  • domain assumption Baseline results in Tables 4-6 were obtained under the same data selection and evaluation protocol
    The paper does not state that baselines were rerun; prior-published numbers may use different splits or backbones.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fully Few-shot Class-incremental Audio Classification Using Multi-level Embedding Extractor and Ridge Regression Classifier." pith.science (2026). https://pith.science/paper/XQMSVTPD

@misc{pith2026250618406,
  author       = {Pith},
  title        = {Pith review of: Fully Few-shot Class-incremental Audio Classification Using Multi-level Embedding Extractor and Ridge Regression Classifier},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XQMSVTPD}},
  note         = {Machine review of arXiv:2506.18406}
}
read the original abstract

In the task of Few-shot Class-incremental Audio Classification (FCAC), training samples of each base class are required to be abundant to train model. However, it is not easy to collect abundant training samples for many base classes due to data scarcity and high collection cost. We discuss a more realistic issue, Fully FCAC (FFCAC), in which training samples of both base and incremental classes are only a few. Furthermore, we propose a FFCAC method using a model which is decoupled into a multi-level embedding extractor and a ridge regression classifier. The embedding extractor consists of an encoder of audio spectrogram Transformer and a fusion module, and is trained in the base session but frozen in all incremental sessions. The classifier is updated continually in each incremental session. Results on three public datasets show that our method exceeds current methods in accuracy, and has advantage over most of them in complexity. The code is at https://github.com/YongjieSi/MAR.

Figures

Figures reproduced from arXiv: 2506.18406 by the authors.

Figure 1
Figure 1. Framework of the proposed method. AST: Audio Spectrogram Transformer; MEE: Multi-level Embedding Extractor; RRC: Ridge Regression Classifier; G and C: the matrices defined by Eq. 6 and 7, respectively; 𝑾௠: classifier weights in session m. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. depicts the structure of the MEE which includes an operation of patch split, an AST encoder and a fusion module. Each LMS is divided into Z sub-spectra along the time and frequency axes with a step size of d. The size of a LMS and a sub-spectrum are Sf ×St and sf ×st, respectively. Sf (or sf) and St (or st) denote the LMS’s (or sub-spectrum’s) dimensions of frequency and time, respectively. Z is computed by 𝑍 = ቔ ௌ೑… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages

  1. [44]

    Few-shot class-incremental audio classification using stochastic classifier,

    Y. Li, et al, “Few-shot class-incremental audio classification using stochastic classifier,” in Proc. of Interspeech, 2023, pp. 4174-4178

    work page 2023

  2. [39]

    Class-incremental grouping network for continual audio-visual learning,

    S. Mo, et al, “Class-incremental grouping network for continual audio-visual learning,” in Proc. of ICCV, 2023, pp. 7754-7764

    work page 2023

  3. [38]

    An efficient incremental learning algorithm for sound classification,

    M.A. Hussain, et al, “An efficient incremental learning algorithm for sound classification,” IEEE MultiMedia, vol. 30, no. 1, pp. 84-90, 2023

    work page 2023

  4. [1]

    Introduction Audio classification (AC) is a task to classify different types of sounds. It is a key module for implementing many audio or video processing tasks, such as acoustic event detection [1]-[5], acoustic scene classification [6]-[9], video analysis [10]-[12], speaker analysis [13]-[15], and keyword spotting [16]-[19]. Many efforts are made on AC....

    work page

  5. [2]

    The proposed MEE can extract discriminative embeddings by fusing features output by multiple blocks of the AST encoder

    We design a MEE which consists of an encoder of Audio Spectrogram Transformer (AST) [40] and a fusion module. The proposed MEE can extract discriminative embeddings by fusing features output by multiple blocks of the AST encoder

    work page

  6. [3]

    As a result, the overfitting issue caused by few training samples is expected to be alleviated

    We design a RRC which can reduce the correlations between various dimensions of embeddings. As a result, the overfitting issue caused by few training samples is expected to be alleviated. The RRC weights are updated by an analytical solution without being retrained in each incremental session

    work page

  7. [4]

    Results show that our method exceeds previous methods in Average Accuracy (AA) and Performance Degradation rate (PD), and has advantage over most of them in complexity

    We propose a FFCAC method using a model which is decoupled into a MEE and a RRC. Results show that our method exceeds previous methods in Average Accuracy (AA) and Performance Degradation rate (PD), and has advantage over most of them in complexity

    work page

  8. [5]

    Method In this section, we describe our method in detail, including problem definition, method framework, multi-level embedding extractor, and ridge regression classifier. 2.1. Problem Definition The FFCAC problem has two types of sessions: base session (Session 0) and incremental session (Sessions 1 to M). M is total number of incremental sessions. The t...

    work page

Show all 58 references
  1. [6]

    Experimental Datasets Table 1 lists detailed information of the LS-100, FSC-89, and NSynth-100

    Experiments 3.1. Experimental Datasets Table 1 lists detailed information of the LS-100, FSC-89, and NSynth-100. The three datasets have been widely adopted for the task of AC and can be obtained from the three websites: https://www.modelscope.cn/datasets/pp199124903/LS-100/su...

  2. [7]

    Based on the experimental results and analyses, we can draw three conclusions

    Conclusions In this work, we propose a FFCAC method using the MEE and RRC. Based on the experimental results and analyses, we can draw three conclusions. First, our method exceeds other methods in accuracy. Second, each main part of our method has contributions to the accuracy...

  3. [8]

    10-34), and Guangdong S&T project (2023A0505050116), and Guangdong provincial key laboratory of human digital twin (2022B1212010004)

    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), and Guangdong S&T project (20...

  4. [9]

    Sound event detection with depthwise separable and dilated convolutions,

    K. Drossos, et al, “Sound event detection with depthwise separable and dilated convolutions,” in Proc. of IJCNN, 2020, pp. 1-7

    2020

  5. [10]

    Domestic activities clustering from audio recordings using convolutional capsule autoencoder network,

    Z. Lin, et al., “Domestic activities clustering from audio recordings using convolutional capsule autoencoder network,” in Proc. of ICASSP, 2021, pp. 835-839

    2021

  6. [11]

    Acoustic event diarization in TV/movie audios using deep embedding and integer linear programming,

    Y. Li, et al, “Acoustic event diarization in TV/movie audios using deep embedding and integer linear programming,” Multimedia Tools Appl., vol. 78, no. 23, pp. 33999-34025, 2019

    2019

  7. [12]

    Using multi-stream hierarchical deep neural network to extract deep audio feature for acoustic events detection,

    Y. Li, et al, “Using multi-stream hierarchical deep neural network to extract deep audio feature for acoustic events detection,” Multimedia Tools Appl., vol. 77, no. 1, pp. 897-916, 2018

    2018

  8. [13]

    Anomalous sound detection using deep audio representation and a BLSTM network for audio surveillance of roads,

    Y. Li, et al, “Anomalous sound detection using deep audio representation and a BLSTM network for audio surveillance of roads,” IEEE Access, vol. 6, pp. 58043-58055, 2018

    2018

  9. [14]

    Acoustic scene classification using aggregation of two-scale deep embeddings,

    H.K. Chon, et al, “Acoustic scene classification using aggregation of two-scale deep embeddings,” in Proc. of IEEE ICCT, 2021, vol. 4, pp. 1341-1345

    2021

  10. [15]

    Acoustic scene classification using deep audio feature and BLSTM network,

    Y. Li, et al, “Acoustic scene classification using deep audio feature and BLSTM network,” in Proc. of ICALIP, 2018, pp. 371-374

    2018

  11. [16]

    Low-complexity acoustic scene classification using deep space separable distillation module and multi-label learning,

    K. Wang, et al, “Low-complexity acoustic scene classification using deep space separable distillation module and multi-label learning,” in Report of DCASE2023 Challenge, 2023, pp. 1-4

    2023

  12. [17]

    Low-complexity acoustic scene classification using parallel attention-convolution network,

    Y. Li, et al, “Low-complexity acoustic scene classification using parallel attention-convolution network,” in Proc. of Interspeech, 2024, pp. 567-571

    2024

  13. [18]

    Violence detection in videos based on fusing visual and audio information,

    W. Pang, et al, “Violence detection in videos based on fusing visual and audio information,” in Proc. of ICASSP, 2021, pp. 2260-2264

    2021

  14. [19]

    Predicting skeleton trajectories using a Skeleton‑Transformer for video anomaly detection,

    W. Pang, et al, “Predicting skeleton trajectories using a Skeleton‑Transformer for video anomaly detection,” Multimedia Systems, 2022, vol. 28, pp. 1481-1494

    2022

  15. [20]

    Detecting video anomalies by jointly utilizing appearance and skeleton information,

    W. Pang, et al, “Detecting video anomalies by jointly utilizing appearance and skeleton information,” Expert Syst. Appl., vol. 246, 2024, Art. no. 123135

    2024

  16. [21]

    Lightweight speaker verification using transformation module with feature grouping and fusion,

    Y. Li, et al, “Lightweight speaker verification using transformation module with feature grouping and fusion,” IEEE/ACM TASLP, vol. 32, pp. 794-806, 2024

    2024

  17. [22]

    Speaker verification using attentive multi-scale convolutional recurrent network,

    Y. Li, et al, “Speaker verification using attentive multi-scale convolutional recurrent network,” Applied Soft Computing, 2022, vol. 126, Article no. 109291, pp. 1-11

    2022

  18. [23]

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

    Y. Li, et al, “Speaker clustering by co-optimizing deep representation learning and cluster estimation,” IEEE TMM, vol. 23, pp. 3377-3387, 2021

    2021

  19. [24]

    Cross-domain few-shot open-set keyword spotting using keyword adaptation and prototype reprojection,

    M. Yang, et al, “Cross-domain few-shot open-set keyword spotting using keyword adaptation and prototype reprojection,” in Proc. of ICASSP, 2025. Accepted for publication

    2025

  20. [25]

    Deep enhancement spotting network for low-complexity keyword spotting in noisy environments,

    Y. Chen, et al, “Deep enhancement spotting network for low-complexity keyword spotting in noisy environments,” in Proc. of ICASSP, 2025. Accepted for publication

    2025

  21. [26]

    A stage match for query-by-example spoken term detection based on structure information of query,

    J. Zhan, et al, “A stage match for query-by-example spoken term detection based on structure information of query,” in Proc. of ICASSP, 2021, pp. 6833-6837

    2021

  22. [27]

    A novel re-weighted CTC loss for data imbalance in speech keyword spotting,

    X. Lan, et al, “A novel re-weighted CTC loss for data imbalance in speech keyword spotting,” Chinese Journal of Electronics, vol. 32, no. 3, pp. 465-473, 2023

    2023

  23. [28]

    Data-centric methods for environmental sound classification with limited labels,

    A.R. Syed, et al, “Data-centric methods for environmental sound classification with limited labels,” IEEE/ACM TASLP, vol. 32, pp. 4288-4297, 2024

    2024

  24. [29]

    ASiT: local-global audio spectrogram vision transformer for event classification,

    S.A.A. Ahmed, et al, “ASiT: local-global audio spectrogram vision transformer for event classification,” IEEE/ACM TASLP, vol. 32, pp. 3684-3693, 2024

    2024

  25. [30]

    Automated data augmentation for audio classification,

    Y. Sun et al., “Automated data augmentation for audio classification,” IEEE/ACM TASLP, vol. 32, pp. 2716-2728, 2024

    2024

  26. [31]

    On the choice of the optimal temporal support for audio classification with pre-trained embeddings,

    A. Quelennec, et al, “On the choice of the optimal temporal support for audio classification with pre-trained embeddings,” in Proc. of ICASSP, 2024, pp. 976-980

    2024

  27. [32]

    Sound event detection via dilated convolutional recurrent neural networks,

    Y. Li, et al, “Sound event detection via dilated convolutional recurrent neural networks,” in Proc. of ICASSP, 2020, pp. 286-290

    2020

  28. [33]

    On the transferability of large-scale self-supervision to few-shot audio classification,

    C. Heggan, et al, “On the transferability of large-scale self-supervision to few-shot audio classification,” in Proc. of ICASSPW, 2024, pp. 515-519

    2024

  29. [34]

    Few-shot anomalous sound detection based on anomaly map estimation using pseudo abnormal data,

    R. Tanaka, et al, “Few-shot anomalous sound detection based on anomaly map estimation using pseudo abnormal data,” in Proc. of ICASSP, 2024, pp. 1391-1395

    2024

  30. [35]

    Learning from taxonomy: multi-label few-shot classification for everyday sound recognition,

    J. Liang, et al, “Learning from taxonomy: multi-label few-shot classification for everyday sound recognition,” in Proc. of ICASSP, 2024, pp. 771-775

    2024

  31. [36]

    Zero- and few-shot sound event localization and detection,

    K. Shimada et al., “Zero- and few-shot sound event localization and detection,” in Proc. of ICASSP, 2024, pp. 636-640

    2024

  32. [37]

    Class-incremental learning for multi- label audio classification,

    M. Mulimani, et al, “Class-incremental learning for multi- label audio classification,” in Proc. of ICASSP, 2024, pp. 916-920

    2024

  33. [40]

    Few shot continual learning for audio classification,

    Y. Wang, et al, “Few shot continual learning for audio classification,” in Proc. of ICASSP, 2021, pp. 321-325

    2021

  34. [41]

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

    Y. Li, et al, “Few-shot class- incremental audio classification using dynamically expanded classifier with self-attention modified prototypes,” IEEE TMM, vol. 26, pp. 1346-1360, 2024

    2024

  35. [42]

    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

    2023

  36. [43]

    Few-shot class-incremental audio classification using adaptively-refined prototypes,

    W. Xie, et al., “Few-shot class-incremental audio classification using adaptively-refined prototypes,” in Proc. of Interspeech, 2023, pp. 301-305

    2023

  37. [45]

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

    Y. Li, et al, “Few-shot class-incremental audio classification with adaptive mitigation of forgetting and overfitting,” IEEE/ACM TASLP, vol. 32, pp. 2297-2311, 2024

    2024

  38. [46]

    Fully few-shot class-incremental audio classification using expandable dual-embedding extractor,

    Y. Si, et al, “Fully few-shot class-incremental audio classification using expandable dual-embedding extractor,” in Proc. of Interspeech, 2024, pp. 4788-4792

    2024

  39. [47]

    Fully few-shot class-incremental audio classification with adaptive improvement of stability and plasticity,

    Y. Si, et al, “Fully few-shot class-incremental audio classification with adaptive improvement of stability and plasticity,” IEEE TASLP, vol. 33, pp. 418-433, 2025

    2025

  40. [48]

    AST: Audio Spectrogram Transformer,

    Y. Gong, et al, “AST: Audio Spectrogram Transformer,” in Proc. of Interspeech, 2021, pp. 571-575

    2021

  41. [49]

    Acoustic scene clustering using joint optimization of deep embedding learning and clustering iteration,

    Y. Li, et al, “Acoustic scene clustering using joint optimization of deep embedding learning and clustering iteration,” IEEE TMM, vol. 22, no. 6, pp. 1385-1394, 2020

    2020

  42. [50]

    Attention is all you need,

    A. Vaswani, et al., “Attention is all you need,” in Proc. of NIPS, 2017, vol. 30, pp. 6000-6010

    2017

  43. [51]

    Continual learning with pre-trained models: a survey,

    D. Zhou, et al, “Continual learning with pre-trained models: a survey,” in Proc. of IJCAI, 2024 pp. 8363-8371

    2024

  44. [52]

    RanPAC: random projections and pre-trained models for continual learning,

    M.D. McDonnell, et al, “RanPAC: random projections and pre-trained models for continual learning,” in Proc. of NeuIPS, 2023, pp. 12022-12053

    2023

  45. [53]

    Pattern recognition and machine learning,

    C.M. Bishop, “Pattern recognition and machine learning,” Springer-Verlag New York, Inc. 2006

    2006

  46. [54]

    Ridge regression: biased estimation for nonorthogonal problems,

    A.E. Hoerl, et al, “Ridge regression: biased estimation for nonorthogonal problems,” Technometrics, vol.12, no.1, pp.55-67,1970

    1970

  47. [55]

    Learning without memorizing,

    P. Dhar, et al, “Learning without memorizing,” in Proc. of CVPR, 2019, pp. 5133-5141

    2019

  48. [56]

    iCaRL: Incremental classifier and representation learning,

    S.-A. Rebuffi, et al, “iCaRL: Incremental classifier and representation learning,” in Proc. of CVPR, 2017, pp. 5533-5542

    2017

  49. [57]

    Few shot incremental learning with continually evolved classifiers,

    C. Zhang, et al, “Few shot incremental learning with continually evolved classifiers,” in Proc. of CVPR, 2021, pp. 12450-12459

    2021

  50. [58]

    Forward compatible few-shot class incremental learning,

    D.-W. Zhou, et al, “Forward compatible few-shot class incremental learning,” in Proc. of CVPR, 2022, pp. 9046-9056

    2022

Pith tools

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