Pith. sign in

REVIEW 3 major objections 4 minor 54 references

Few-Shot Class-Incremental Audio Classification Using Pseudo-Incrementally Trained Embedding Learner and Continually Updated Stochastic Classifier

T0 review · 3 major / 4 minor · reviewed 2026-07-08 · grok-4.5

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

Pith's one-line read A frozen embedding learner plus a continually updated stochastic classifier identifies few-shot incremental audio classes without forgetting base ones

desk verdict Practical FCAC recipe that freezes a residual embedding after pseudo-incremental base training and keeps a mean–variance classifier plastic; useful subfield work if the frozen-embedding premise holds. read the letter →

arxiv 2607.05953 v1 pith:7R7RCEXY submitted 2026-07-07 eess.AS

Yanxiong Li , Wenchang Cao , Jiaxin Tan , Qianqian Li , Guoqing Chen This is my paper · ORCID
classification eess.AS
keywords few-shotclass-incrementallearningaudioclassificationembeddinglearnerstochasticclassifiercatastrophicforgettingdataaugmentationresidualCNNcontinual
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

Few-shot class-incremental audio classification asks a model to keep recognizing many base classes while adding new classes from only a handful of labeled examples. This paper claims the problem can be solved by decoupling the model into two independent pieces: an embedding learner that is trained once in the base session and then frozen, and a stochastic classifier that is updated in every session. The embedding learner is a residual convolutional network given extra representation capacity for future classes by a pseudo-incremental training strategy that uses data augmentation during the base session alone. The classifier represents each class as a mean vector and a variance vector; both are continually refreshed so the model can adapt to new classes without touching the frozen backbone. On three standard few-shot incremental audio benchmarks the method exceeds prior approaches in accuracy while remaining lighter than most of them. If the decoupling works, practitioners can add new sound classes cheaply without catastrophic forgetting or full retraining.

What carries the argument

The pseudo-incrementally trained embedding learner (PITEL) plus continually updated stochastic classifier (CUSC). PITEL is the residual-CNN backbone trained once with augmentation-driven pseudo-incremental sessions so it already spans future class distributions; CUSC stores each class as a learnable mean and variance vector that is updated every incremental session while the backbone stays frozen.

What would settle it

On a held-out incremental session whose acoustic distribution lies well outside the base-session augmentation manifold, measure whether accuracy on the new classes collapses while base-class accuracy remains high; collapse would show the frozen embedding cannot represent those classes.

Watch

Extended reading notes

Core claim

A residual-CNN embedding learner that is pseudo-incrementally trained with data augmentation in the base session and then frozen, coupled with a continually updated stochastic classifier of per-class mean and variance vectors, can progressively recognize few-shot incremental audio classes while retaining base-class knowledge, outperforming comparison methods in accuracy on FSC-89, NSynth-100 and LS-100 at lower complexity than most of them.

Load-bearing premise

That representation capacity for all future incremental classes can be adequately pre-built in the base session alone via data-augmentation-based pseudo-incremental training, so freezing the embedding thereafter does not cap plasticity.

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, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a few-shot class-incremental audio classification (FCAC) method that decouples a residual-CNN embedding learner from a stochastic classifier whose classes are represented by mean and variance vectors. After base-session training, the embedding is frozen to preserve base-class knowledge; plasticity is obtained by continually updating the stochastic classifier on few-shot incremental samples. To give the frozen embedding capacity for future classes, the authors introduce a base-session pseudo-incremental training strategy that uses data augmentation to simulate incremental sessions. The abstract claims higher accuracy than comparison methods on FSC-89, NSynth-100, and LS-100, together with lower complexity than most baselines, and points to a public code release.

Significance. If the empirical claims hold under standard FCAC protocols, the work offers a practical and relatively simple stability–plasticity recipe for audio: freeze a residual-CNN backbone after a carefully designed base session, and adapt only a lightweight stochastic classifier thereafter. Lower complexity than most competitors and a public code link are concrete strengths. The contribution is primarily empirical and architectural rather than theoretical; its value to the community rests on whether the frozen-embedding premise is shown to generalize to truly held-out incremental classes and on the quality and completeness of the experimental comparison.

major comments (3)
  1. The central stability–plasticity claim rests on a load-bearing premise stated in the abstract: that base-session data-augmentation pseudo-increments can pre-build embedding capacity for all future few-shot classes so that freezing the residual-CNN thereafter does not impose a hard ceiling on plasticity. Augmentation expands coverage only within the acoustic manifold of the base classes. If held-out incremental classes on FSC-89, NSynth-100, or LS-100 occupy a different region of feature space, classifier-only updates of mean/variance vectors cannot recover. This premise must be stress-tested with (i) an ablation that removes or degrades the pseudo-incremental strategy, (ii) a frozen vs. unfrozen embedding comparison under matched budgets, and (iii) a feature-space analysis (e.g., nearest-class or silhouette statistics) showing that real incremental classes remain separable under the froz
  2. The abstract asserts superior accuracy and lower complexity on three datasets but, as presented, does not specify session protocols, class splits, number of base vs. incremental classes, shots per class, number of incremental sessions, or how baselines were re-implemented. These details are load-bearing for the cross-method claim. The manuscript must report full session schedules, random seeds or multiple runs with error bars, and a complexity table (parameters, FLOPs or wall-clock) that makes the “lower complexity than most” claim checkable. If any baseline was not re-run under the same splits and augmentation budget, that must be stated and justified.
  3. The stochastic classifier is described only at the level of “mean and variance vectors per class.” How these statistics are estimated from few-shot samples (sample moments, Bayesian update, EMA, temperature, covariance structure—diagonal only?), how class scores are formed at inference (Mahalanobis, sampling, closed-form Gaussian likelihood), and whether base-class statistics are frozen or also refreshed when new classes arrive are all unspecified in the abstract-level method statement. These choices directly affect both plasticity and catastrophic forgetting; they need precise definitions and, where relevant, ablations, because they are part of the continually updated module that is claimed to deliver the accuracy gains.
minor comments (4)
  1. Define FCAC and the three dataset names (FSC-89, NSynth-100, LS-100) with citations and brief construction notes (source corpus, number of classes, base/incremental split) on first use so readers outside the immediate sub-community can follow.
  2. Clarify the residual-CNN architecture (depth, input representation—log-mel, waveform, etc.—and any pretraining) when the embedding learner is introduced; “backbone of a residual convolutional network” is too vague for reproducibility even with a code link.
  3. State explicitly whether the pseudo-incremental augmentation schedule (types, magnitudes, number of pseudo-sessions) is fixed across datasets or tuned per dataset; if tuned, report the search space to avoid hidden free parameters.
  4. The phrase “increasingly identify incremental classes without forgetting base classes” should be tied to a concrete metric suite (e.g., average incremental accuracy, performance dropping rate, base-class retention) rather than left as a qualitative claim.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The three major comments correctly identify places where the manuscript must make its stability–plasticity premise, experimental protocol, and stochastic-classifier design fully checkable. We agree that these points are load-bearing for the empirical claims and will revise accordingly: (i) add the requested ablations and a feature-space analysis of real incremental classes under the frozen embedding; (ii) report complete session schedules, multi-run statistics with error bars, and an explicit complexity table; and (iii) give precise definitions of how class mean/variance statistics are estimated, how scores are formed at inference, and how base-class statistics are treated when new classes arrive. We believe these revisions will place the contribution on a firmer empirical footing without changing the core method.

read point-by-point responses

  1. Referee: The central stability–plasticity claim rests on the premise that base-session data-augmentation pseudo-increments pre-build embedding capacity for future few-shot classes so freezing the residual-CNN does not impose a hard ceiling. Augmentation expands coverage only within the base-class manifold. This premise must be stress-tested with (i) an ablation that removes/degrades the pseudo-incremental strategy, (ii) a frozen vs. unfrozen embedding comparison under matched budgets, and (iii) a feature-space analysis showing real incremental classes remain separable under the frozen embedding.

    Authors: We agree that this premise is load-bearing and that the current manuscript does not stress-test it sufficiently. Conceptually, the pseudo-incremental base-session strategy is not intended to invent out-of-manifold classes; it forces the residual-CNN, while still trainable, to repeatedly form decision structure for held-out base subsets treated as “new” sessions under augmentation, so that the frozen embedding later retains usable geometry for true incremental classes. That rationale does not replace empirical verification. In the revision we will: (i) ablate or degrade the pseudo-incremental strategy (no pseudo-sessions; weaker augmentation; fewer pseudo-sessions) and report the accuracy drop on all three datasets; (ii) compare frozen vs. unfrozen embeddings under matched parameter/update budgets and the same few-shot incremental protocol, so any plasticity ceiling is quantified rather than assumed; and (iii) add a feature-space analysis (nearest-class mean distances and silhouette-type statistics, plus qualitative embeddings) of real incremental classes under the frozen backbone. If incremental classes prove poorly separated, we will state that limitation explicitly. These additions directly address the referee’s three requested tests. revision: yes

  2. Referee: The abstract asserts superior accuracy and lower complexity on three datasets but does not specify session protocols, class splits, number of base vs. incremental classes, shots per class, number of incremental sessions, or how baselines were re-implemented. The manuscript must report full session schedules, random seeds or multiple runs with error bars, and a complexity table (parameters, FLOPs or wall-clock). If any baseline was not re-run under the same splits and augmentation budget, that must be stated and justified.

    Authors: The referee is right that the cross-method claim is only as strong as the protocol and complexity accounting. The experimental section already states the datasets and high-level FCAC setting, but it does not present the information in a form that makes every comparison fully checkable. In the revision we will: (1) give a complete session schedule table for FSC-89, NSynth-100, and LS-100 (base-class count, incremental classes per session, number of sessions, shots per class, and evaluation protocol); (2) report means and standard deviations over multiple random class-order/seeds rather than single-run point estimates; (3) add a complexity table listing trainable parameters after the base session, parameters updated in each incremental session, and FLOPs or measured wall-clock for embedding forward pass and classifier update, so the “lower complexity than most baselines” claim can be verified line by line; and (4) state explicitly, for each baseline, whether it was re-run under our splits and augmentation budget or taken from prior reports, and justify any non-re-run case. We will also align the abstract wording with these tables so the superiority claim is scoped to the reported protocol. revision: yes

  3. Referee: The stochastic classifier is described only as “mean and variance vectors per class.” How these statistics are estimated from few-shot samples (sample moments, Bayesian update, EMA, temperature, covariance structure—diagonal only?), how class scores are formed at inference (Mahalanobis, sampling, closed-form Gaussian likelihood), and whether base-class statistics are frozen or also refreshed when new classes arrive are unspecified. These choices affect plasticity and forgetting and need precise definitions and, where relevant, ablations.

    Authors: We agree that the continually updated module is under-specified at the level of the current method statement, and that those choices are part of the claimed accuracy gains. In the revision we will define the classifier formally: each class is a diagonal Gaussian in the embedding space, parameterized by a mean vector and a per-dimension variance vector; statistics for a new class are estimated from the few-shot embeddings by sample moments (with a stated variance floor/prior to stabilize one- or few-shot estimates); inference scores are closed-form Gaussian log-likelihoods (equivalently, a diagonal Mahalanobis distance plus log-determinant term), without Monte Carlo sampling at test time; and base-class mean/variance vectors are retained from the base session and not refreshed on incremental data, while only newly arrived classes receive newly estimated statistics—this is the mechanism intended to limit catastrophic forgetting of base classes while allowing plasticity for new ones. We will add the corresponding equations, implementation details (temperature/scaling if used), and a short ablation on variance estimation and on whether base-class statistics are left frozen vs. lightly updated. These definitions will appear in the main method section so the continually updated module is fully reproducible from the paper together with the public code. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: empirical FCAC method with independent modules and external benchmarks; no derivation reduces a prediction to its inputs by construction.

full rationale

This is an empirical methods paper proposing PITEL-CUSC for few-shot class-incremental audio classification. The claimed gains rest on (1) freezing a residual-CNN embedding after base-session pseudo-incremental training via data augmentation and (2) continually updating a stochastic classifier (per-class mean/variance). Neither step is definitionally equivalent to its inputs: pseudo-incremental training is a design choice whose adequacy is tested on held-out incremental classes of FSC-89, NSynth-100, and LS-100 against external comparison methods; accuracy and complexity are measured, not forced by a fitted constant renamed as prediction. There is no uniqueness theorem imported from the authors, no ansatz smuggled via self-citation that forbids alternatives, and no self-definitional equation (e.g., a scale fitted from the same ratio later “predicted”). Self-citations, if any, are ordinary prior-work context and not load-bearing for the central empirical claim. The reader’s residual concern—that base-session augmentation may not pre-build capacity for truly novel acoustic manifolds—is a correctness/assumption risk, not circularity under the stated criteria. Score 0 is therefore the honest finding: the derivation chain is self-contained against external benchmarks and does not reduce by construction.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

Abstract-only audit. The method rests on standard deep-learning and few-shot/continual-learning assumptions rather than new physical entities. Free parameters (network width/depth, learning rates, augmentation schedule, variance initialization, number of base vs incremental classes, shots per class) are implied but not numerically reported here. No invented particles or forces. Main domain assumptions: residual CNN embeddings transfer across audio classes; class-conditional mean/variance vectors suffice as a classifier under few-shot updates; base-session augmentation can stand in for future class diversity.

free parameters (3)
  • Base-session augmentation / pseudo-incremental schedule
    The abstract states the embedding is pseudo-incrementally trained with data augmentation in the base session; the specific transforms, strengths, and simulated-class protocol are free design choices that directly affect claimed plasticity after freeze.
  • Stochastic classifier mean and variance vectors (per class)
    Each class is represented by a mean vector and a variance vector that are fit/updated from the available labeled samples; these are data-dependent parameters of the classifier head.
  • Residual CNN embedding hyperparameters
    Architecture depth/width, training epochs, optimizer settings, and feature dimension are standard free hyperparameters of the backbone; values are not given in the abstract.
assumptions (4)
  • domain assumption Freezing the embedding after the base session preserves base-class knowledge (stability) while still allowing incremental recognition via classifier updates alone.
    Core design premise of the decoupled model stated in the abstract; not proved, only motivated and to be tested empirically.
  • ad hoc to paper Data augmentation in the base session can pseudo-simulate incremental classes well enough that the frozen embedding generalizes to real future few-shot classes.
    The pseudo-incremental training strategy is the paper’s specific plasticity mechanism for the frozen backbone; its adequacy is an unproved modeling choice.
  • domain assumption A class can be adequately represented for few-shot audio classification by a mean vector and a variance vector in embedding space.
    Stochastic/prototype-style classifier assumption common in few-shot learning; adopted here without independent derivation in the abstract.
  • domain assumption Standard residual convolutional networks are suitable audio embedding backbones for the FCAC setting.
    Background architectural choice stated as the embedding learner backbone.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Few-Shot Class-Incremental Audio Classification Using Pseudo-Incrementally Trained Embedding Learner and Continually Updated Stochastic Classifier." pith.science (2026). https://pith.science/paper/7R7RCEXY

@misc{pith2026260705953,
  author       = {Pith},
  title        = {Pith review of: Few-Shot Class-Incremental Audio Classification Using Pseudo-Incrementally Trained Embedding Learner and Continually Updated Stochastic Classifier},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7R7RCEXY}},
  note         = {Machine review of arXiv:2607.05953}
}
read the original abstract

Few-shot Class-incremental Audio Classification (FCAC) aims to progressively recognize incremental classes with few tagged samples and meanwhile memorize base classes. To achieve satisfactory FCAC performance, the model needs to have high stability (memorizing base classes) and strong plasticity (adapting to incremental classes). In this work, we design a model which can be decoupled into two independent modules, namely an embedding learner and a stochastic classifier. The former is the backbone of a residual convolutional network, while the latter is composed of distributions and each distribution consists of a mean vector and a variance vector for representing one class. After being trained in the base session, the embedding learner is not updated in each incremental session and thus can memorize the knowledge of base classes. To make the embedding learner possess strong representation ability for incremental classes, we propose a strategy to pseudo-incrementally train the embedding learner using data augmentation in the base session. On the other hand, the stochastic classifier is continually updated in each incremental session and thus can adapt to incremental classes. Our model which consists of a pseudo-incrementally trained embedding learner and a continually updated stochastic classifier can increasingly identify incremental classes without forgetting base classes. Three datasets (FSC-89, NSynth-100 and LS-100) are used to verify the effectiveness of our method. Experiments show that our method exceeds the comparison methods in accuracy, and has lower complexity than most of the comparison methods. The code is at https://github.com/vinceasvp/PITEL-CUSC.

Figures

Figures reproduced from arXiv: 2607.05953 by the authors.

Figure 1
Figure 1. The framework of few-shot class-incremental audio class [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Framework for our method. Model 𝜽଴ is first trained in supervised learning way with abundant samples and is then updated in few-shot learning way (pseudo-incremental training) in base session. Embedding learner 𝑬଴ is frozen and stochastic classifier 𝑪௠ is updated in few-shot learning way in the mth incremental session. 𝑼଴ = {𝝁଴,௡}, 1 ≤ n ≤ N0, and 𝝁଴,௡ denotes the mean vector of embeddings of the nth class in base s… view at source ↗
Figure 3
Figure 3. Average pooling Log Mel-spectrum ResNet module 4 ResNet module 1 3×3 Conv, 64 Embedding ResNet module 2 ResNet module 3 ResNet module 4 ResNet module 3 ResNet module 2 ResNet module 1 Fully-connected, 512 ReLU 3×3 Conv, 64 3×3 Conv, 64 ReLU ReLU 3×3 Conv, 256 3×3 Conv, 256 ReLU ReLU 3×3 Conv, 128 3×3 Conv, 128 ReLU : Element-wise summation ReLU 3×3 Conv, 512 3×3 Conv, 512 ReLU Audio sample [PITH_FULL_IMAGE:figures/… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: depicts the difference between a stochastic classifier and a deterministic classifier. The area inside the dashed circle in [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Results of statistical significance test for all method [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]
Figure 6
Figure 6. Figure 6: Influence of the NP values on the AA scores obtained by [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: (a) Accuracy scores obtained by our method in the last [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: (a), (b) and (c), respectively. When α equals 0.4, 3.0 and 4.0, our method obtains the highest AA scores on the FSC-89, NSynth-100 and LS-100, respectively. When α changes, the variation in AA scores on the three datasets is less than 0.50% (42.00% - 41.50%). Hence, ou…
Figure 9
Figure 9. Figure 9: The impact of M′ on the AA scores obtained by our method on the (a) FSC-89; (b) NSynth-100; and (c) LS-100. In the fifth extended experiment, we discuss feature space adaptation through plotting the embeddings before and after performing the PITS. Without losing genera…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 54 canonical work pages

  1. [1]

    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 IEEE ICASSP , 2021, pp. 835-839

    work page 2021

  2. [2]

    Anomalous sou nd detection using deep audio representation and a BLSTM network f or audio surveillance of roads,

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

    work page 2018

  3. [3]

    Sound event de tection via dilated convolutional recurrent neural networks,

    Y. Li, M. Liu, K. Drossos, and T. Virtanen, “Sound event de tection via dilated convolutional recurrent neural networks,” in Proc. of IEEE ICASSP, 2020, pp. 286-290. [ 5 ] J . T a n , a n d Y . L i , “ L o w - c o m plexity acoustic scene classific ation using blueprint separable convolution and knowledge distillation,” in Technical Report of DCASE2023 Ch...

    work page 2020

  4. [4]

    Low-c omplexity acoustic scene classification using data augmentation and light weight ResNet,

    Y. Li, W. Cao, W. Xie, Q. Huang, W. Pang, and Q. He, “Low-c omplexity acoustic scene classification using data augmentation and light weight ResNet,” in Proc. of IEEE ICSP, 2022, pp. 41-45

    work page 2022

  5. [5]

    Acoustic scene classific ation using deep CNNs with time-frequency representations,

    W. Xie, Q. He, H. Yan, and Y. Li, “Acoustic scene classific ation using deep CNNs with time-frequency representations,” in Proc. of IEEE ICCT, 2021, vol. 4, pp. 1325-1329

    work page 2021

  6. [6]

    Low-compl exity acoustic scene classification using parallel attention-convolution network

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

    work page 2024

  7. [7]

    Detecting video anoma lies by jointly utilizing appearance a nd skeleton information,

    W. Pang, Q. He, Y. Li, and N. Ahmed, “Detecting video anoma lies by jointly utilizing appearance a nd skeleton information,” Expert Systems with Applications, vol. 246, pp. 1-12, 2024

    work page 2024

  8. [8]

    Audiovisual dependency attention for vio lence detection in videos,

    W. Pang, et al., “Audiovisual dependency attention for vio lence detection in videos,” IEEE TMM, vol. 25, pp. 4922-4932, 2023

    work page 2023

Show all 54 references
  1. [9]

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

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

    2021

  2. [10]

    Interrelate training and clustering for online speaker diarization,

    Y. Chen, et al., “Interrelate training and clustering for online speaker diarization,” IEEE/ACM TASLP, vol. 32, pp. 1352-1364, 2024

    2024

  3. [11]

    Lightweight s p e a k e r verification using transformation module with feature grouping and fusion,

    Y. Li, Z. Jiang, Q. Huang, W. Cao, and J. Li, “Lightweight s p e a k e r verification using transformation module with feature grouping and fusion,” IEEE/ACM TASLP, vol. 32, pp. 794-806, 2024

    2024

  4. [12]

    Few-shot spe aker identification using lightweight prototypical network with feat ure grouping and interaction,

    Y. Li, H. Chen, W. Cao, Q. Huang, and Q. He, “Few-shot spe aker identification using lightweight prototypical network with feat ure grouping and interaction,” IEEE TMM, vol. 25, pp. 9241-9253, 2023

    2023

  5. [13]

    Deep recurrent neural networks for audio classification in construction sites,

    M. Scarpiniti, et al., “Deep recurrent neural networks for audio classification in construction sites,” in Proc. of EUSIPCO, 2021, pp. 810-814

    2021

  6. [14]

    Pyramidal temporal pooling with discriminative mapping for audio classification,

    L. Zhang, Z. Shi, and J. Han, “Pyramidal temporal pooling with discriminative mapping for audio classification,” IEEE/ACM TASLP, vol. 28, pp. 770-784, 2020

    2020

  7. [15]

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

    A. Quelennec, M. Olvera, G. Peeters, and S. Essid, “On the choice of the optimal temporal support for audio classification with pre-trai ned embeddings,” in Proc. of IEEE ICASSP, 2024, pp. 976-980

    2024

  8. [16]

    Multi-level graph learning for audio even t classification and human-perceived annoyance rating prediction,

    Y. Hou, et al., “Multi-level graph learning for audio even t classification and human-perceived annoyance rating prediction,” in Proc. of IEEE ICASSP, 2024, pp. 716-720

    2024

  9. [17]

    Max-AST: combining convolution, local and global self-attentions for aud io event classification,

    T. Alex, S. Ahmed, A. Mustafa, M. Awais, and P. J. Jackson, “Max-AST: combining convolution, local and global self-attentions for aud io event classification,” in Proc. of IEEE ICASSP, 2024, pp. 1061-1065

    2024

  10. [18]

    Active learning for sound event classification using bayesian neural networks with gaussian variational posterior,

    S. Shishkin, D. Hollosi, S. Goetze, and S. Doclo, “Active learning for sound event classification using bayesian neural networks with gaussian variational posterior,” in Proc. of IEEE ICASSP, 2024, pp. 896-900

    2024

  11. [19]

    Automated data augmentation for audio cla ssification,

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

    2024

  12. [20]

    A weig hted binary cross-entropy for sound event representation learning and few-s hot classification,

    Z. Bai, C. Pan, G. Chen, J. Chenk, and J. Benesty, “A weig hted binary cross-entropy for sound event representation learning and few-s hot classification,” in Proc. of APSIPA ASC, 2023, pp. 1069-1074

    2023

  13. [21]

    Frame-le vel embedding learning for few- shot bioacoustic event detection,

    X. Zhang, et al., “Frame-le vel embedding learning for few- shot bioacoustic event detection,” in Proc. of IEEE ICME, 2023, pp. 750-755

    2023

  14. [22]

    Regularized cont rastive pre- training for few-shot bioacoustic sound detection,

    I. Moummad, N. Farrugia, and R. Serizel, “Regularized cont rastive pre- training for few-shot bioacoustic sound detection,” in Proc. of IEEE ICASSP, 2024, pp. 1436-1440

    2024

  15. [23]

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

    J. Liang, H. Phan, and E. Benetos, “Learning from taxonomy : multi-label few-shot classification for ev eryday sound recognition,” in Proc. of IEEE ICASSP, 2024, pp. 771-775

    2024

  16. [24]

    Continual learning of new sound classes using generative replay,

    Z. Wang, et al., “Continual learning of new sound classes using generative replay,” in Proc. of WASPAA, 2019, pp. 308-312

    2019

  17. [25]

    Class-incremental learning f or multi- label audio classification,

    M. Mulimani, and A. Mesaro s, “Class-incremental learning f or multi- label audio classification,” in Proc. of IEEE ICASSP, 2024, pp. 916-920

    2024

  18. [26]

    An Efficient Incremental Learning A lgorithm for Sound Classification,

    M.A. Hussain, et al., “An Efficient Incremental Learning A lgorithm for Sound Classification,” IEEE MultiMedia, vol. 30, no. 1, pp. 84-90, 2023

    2023

  19. [27]

    Few-shot continual learning for audio cl assification,

    Y. Wang, et al., “Few-shot continual learning for audio cl assification,” in Proc. of IEEE ICASSP, 2021, pp. 321-325. [ 3 0 ] Y . L i , e t a l . , “ F e w - s h o t c l a s s - i n c r e m e n t a l a u d i o c l a s s i f i cation using dynamically expanded classifier with s...

    2021

  20. [28]

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

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

    2023

  21. [29]

    Few-shot class-in cremental audio classification with adaptive mi tigation of forgetting and overf itting,

    Y. Li, J. Li, Y. Si, J. Tan, and Q. He, “Few-shot class-in cremental audio classification with adaptive mi tigation of forgetting and overf itting,” IEEE/ACM TASLP, vol. 32, pp. 2297-2311, 2024

    2024

  22. [30]

    Stochasti c-YOLO: Efficient probabilistic object de tection under dataset shifts,

    T. Azevedo, R.d. Jong, M. M attina, and P. Maji, “Stochasti c-YOLO: Efficient probabilistic object de tection under dataset shifts,” i n Proc. of Workshop on Machine Learning for Autonomous Driving at the NeurIPS , 2020, pp. 1-9

    2020

  23. [31]

    A deep sto chastic adaptive Fourier decomposition ne twork for hyperspectral image classification,

    C. Cheng, L. Zhang, H. Li, L. Dai, and W. Cui, “A deep sto chastic adaptive Fourier decomposition ne twork for hyperspectral image classification,” IEEE TIP, vol. 33, pp. 1080-1094, 2024

    2024

  24. [32]

    Mixup: be yond empirical risk minimization,

    H. Zhang, M. Cisse, Y.N. Dauphin, D. Lopez-Paz, “Mixup: be yond empirical risk minimization,” in Proc. of ICLR, 2018, pp. 1-13

    2018

  25. [33]

    Stochastic c lassifiers for unsupervised do main adaptation,

    Z. Lu, et al., “Stochastic c lassifiers for unsupervised do main adaptation,” in Proc. of IEEE/CVF CVPR, 2020, pp. 9108-9117. [ 3 8 ] Y . L i , e t a l . , “ F e w - s h o t c l a s s - i n c r e m e n t a l a u d i o c l a s s i f i cation using stochastic classifier,” in Pro...

    2020

  26. [34]

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

    Y. Li, M. Liu, W. Wang, Y. Zhang, and Q. He, “Acoustic sce ne clustering using joint optimization of deep embedding learning and clustering iteration,” IEEE TMM, vol. 22, no. 6, pp. 1385-1394, 2020

    2020

  27. [35]

    Domestic activity clustering from audio vi a depthwise separable convolutional autoencoder network,

    Y. Li, et al., “Domestic activity clustering from audio vi a depthwise separable convolutional autoencoder network,” in Proc. IEEE 24th Int. Workshop Multimedia Signal Process., 2022, pp. 1-6

    2022

  28. [36]

    Deep mutual attention network for acousti c scene classification,

    W. Xie, et al., “Deep mutual attention network for acousti c scene classification,” Digital Signal Processing, vol. 123, 2022, Art. no. 103450

    2022

  29. [37]

    Speaker ver ification using attentive multi -scale convolutional recurrent network,

    Y. Li, et al., “Speaker ver ification using attentive multi -scale convolutional recurrent network,” Appl. Soft Comput., vol. 126, 2022, Art. no. 109291

    2022

  30. [38]

    Deep residual learni ng for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learni ng for image recognition,” in Proc. of IEEE/CVF CVPR, 2016, pp. 770-778

    2016

  31. [39]

    QARV: Quantization-aware ResNet VAE for lossy image compression,

    Z. Duan, et al., “QARV: Quantization-aware ResNet VAE for lossy image compression,” IEEE TPAMI, vol. 46, no. 1, pp. 436-450, 2024

    2024

  32. [40]

    A four-stage data augmentation approach to ResNet- Conformer based acoustic modeling for sound event localization and detection,

    Q. Wang, et al., “A four-stage data augmentation approach to ResNet- Conformer based acoustic modeling for sound event localization and detection,” IEEE/ACM TASLP, vol. 31, pp. 1251-1264, 2023

    2023

  33. [41]

    Bayesian learning for neural networks: an algorithmic survey,

    M. Magris, and A. Iosifdis, “Bayesian learning for neural networks: an algorithmic survey,” Artificial Intelligence Review , v o l . 5 6 , n o . 1 0 , p p . 11773-11823, 2023

    2023

  34. [42]

    Hands-on Bayesian neural networks—A tutorial for deep learning users,

    L.V. Jospin, et al., “Hands-on Bayesian neural networks—A tutorial for deep learning users,” IEEE Computational Intelligence Magazine, vol. 17, no. 2, pp. 29-48, 2022

    2022

  35. [43]

    Who calls the shots? Rethinking few-shot learning for audio,

    Y. Wang, N. J. Bryan, J. Salamon, M. Cartwright, and J. P. Bello, “Who calls the shots? Rethinking few-shot learning for audio,” in Proc. of IEEE WASPAA, 2021, pp. 36-40

    2021

  36. [44]

    Neural audio synthesis of musical notes with Wave Net autoencoders,

    J. Engel, C. Resnick, A. Roberts, S. Dieleman, M. Norouzi, D. Eck, and K. Simonyan, “Neural audio synthesis of musical notes with Wave Net autoencoders,” in Proc. of ICML, 2017, pp. 1068-1077

    2017

  37. [45]

    Libris peech: An ASR corpus based on public domain audio books,

    V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Libris peech: An ASR corpus based on public domain audio books,” in Proc. of IEEE ICASSP, 2015, pp. 5206-5210

    2015

  38. [46]

    Learning without memorizing,

    P. Dhar, R.V. Singh, K.-C. Peng, Z. Wu, and R. Chellappa, “Learning without memorizing,” in Proc. IEEE/CVF CVPR, 2019, pp. 5138-5146. 15

    2019

  39. [47]

    iCaRL: Incremental classifier and r epresentation learning,

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

    2017

  40. [48]

    Few-shot incremental learning with continually evolved classifiers,

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

    2021

  41. [49]

    Forward compatible few-shot class-incr emental learning,

    D.W. Zhou, et al., “Forward compatible few-shot class-incr emental learning,” in Proc. of IEEE/CVF CVPR, 2022, pp. 9036-9046

    2022

  42. [50]

    Margin-based few-shot class-incremental l earning with class-level overfitting mitigation,

    Y. Zou, et al., “Margin-based few-shot class-incremental l earning with class-level overfitting mitigation,” in Proc. of NeurIPS, 2022, pp. 1-13

    2022

  43. [51]

    Learnable distribution calibration for fe w-shot class- incremental learning,

    B. Liu, et al., “Learnable distribution calibration for fe w-shot class- incremental learning,” IEEE TPAMI , vol. 45, no. 10, pp. 12699-12706, 2023

    2023

  44. [52]

    Dynamic support network for few-shot cla ss incremental learning,

    B. Yang, et al., “Dynamic support network for few-shot cla ss incremental learning,” IEEE TPAMI, vol. 45, no. 3, pp. 2945-2951, 2023

    2023

  45. [53]

    Statistical comparisons of classifiers over mu ltiple data sets,

    J. Demšar, “Statistical comparisons of classifiers over mu ltiple data sets,” Journal of Machine Learning Research, vol.7, pp. 1-30, 2006

    2006

  46. [54]

    Visualizing data using T-SN E,

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

    2008

Pith tools

Reviewed July 8, 2026 · model on record in the stance chip above.