Pith. sign in

REVIEW 4 major objections 4 minor 28 references

Online incremental learning for audio classification using a pretrained audio model

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read A frozen pretrained audio model can keep learning new sound classes and new recording domains online, in a single forward pass per task, with minimal forgetting.

desk verdict A clean, honest RanPAC adaptation to audio; the key weakness is Q chosen on the test tasks and the absence of a RanPAC baseline. read the letter →

arxiv 2508.20732 v1 pith:W463TRFX submitted 2025-08-28 eess.AS

classification eess.AS
keywords onlineincrementallearningclass-incrementaldomain-incrementalaudioclassificationpretrainedmodelrandomprojectionsridgeregressioncatastrophicforgetting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a frozen pretrained audio model can serve as a permanent feature extractor for online incremental learning: when new sound classes or new recording domains arrive, the model adapts in a single forward pass through the new data and forgets little of what it learned before. The key move is to insert a random, frozen projection layer with ReLU between the pretrained embeddings and the classifier, then update a Gram matrix and a class-prototype matrix and solve a ridge-regression readout. On ESC-50 class-incremental learning the method reaches 93.4% average accuracy, and on TAU Urban Acoustic Scenes 2019 domain-incremental learning it reaches 61.4%, beating every baseline the paper compares against, including joint linear probes that are allowed to see all previous data. If this holds, incremental audio classification no longer needs base-task training, replay buffers, task IDs, or multi-epoch fine-tuning when new classes or domains appear.

What carries the argument

The central mechanism is the RanPAC recipe adapted to audio: a random nonlinear expansion followed by a ridge-regression readout. Frozen 2048-dimensional pretrained embeddings are projected by a fixed random matrix W into Q = 8192 dimensions, passed through ReLU, and accumulated into a Q×Q Gram matrix and a Q×C class-prototype matrix; inference uses Wo = (G + λI)^−1 K. The random expansion improves linear separability of classes, while the ridge inversion decorrelates the class prototypes so that updating the readout on new tasks does not overwrite the classification structure of old tasks.

What would settle it

Run the identical online protocol on a held-out sequence of sound classes or acoustic domains that are semantically far from AudioSet, such as medical or marine sounds, keeping Q = 8192 and the per-task λ search unchanged. If single-pass accuracy drops to the level of a plain linear probe, the claim that the random expansion and decorrelation are what prevent forgetting would be refuted.

Watch

Extended reading notes

Core claim

The paper demonstrates that a fixed pretrained audio encoder can be turned into an online continual learner without any base-task training. Features from the frozen PANNs CNN14 model are projected through a random matrix into 8192 dimensions, passed through ReLU, and accumulated into a Gram matrix G and an unnormalized class-prototype matrix K; the classifier weights are then computed as Wo = (G + λI)^−1 K, with λ chosen per task on a validation split. This single-pass update yields 93.4% final average accuracy with 2.5% forgetting on the 50-class ESC-50 class-incremental setup, and 61.4% accuracy with 2.0% forgetting on the 9-city TAU Urban Acoustic Scenes 2019 domain-incremental setup. The

Load-bearing premise

The whole method bets that PANNs CNN14's frozen 2048-dimensional embeddings remain informative for any sound class and domain that arrives later, and that the projection size and ridge penalty chosen on the same benchmark tasks are not silently inflating the reported accuracy.

Editorial extensions

If this is right

  • One unified online framework handles both class-incremental and domain-incremental audio learning without task IDs, replay buffers, or base-task training.
  • Adapting to a new task requires only a single forward pass through its training samples, so deployment on streaming or resource-constrained devices becomes practical.
  • The reported accuracy exceeds even a joint linear probe that trains on all tasks together, suggesting that the nonlinear random expansion plus decorrelation recovers useful structure from frozen features.
  • Forgetting stays around 2 percentage points in both setups, indicating that stability-plasticity balance is achieved without explicit memory of past data.
  • Trainable parameters are limited to the Q×C readout plus the accumulated Gram matrix, far fewer than the 80.8M parameters of the frozen CNN14 backbone.

Reading between the lines

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

  • The method's success depends on how well PANNs CNN14 embeddings separate sound classes that were not part of its training distribution; a natural test is to swap in other pretrained audio encoders and see whether accuracy tracks the quality of their embeddings.
  • The authors chose Q = 8192 and the per-task λ search using validation splits on the same ESC-50 and TAU tasks; a stricter evaluation would fix these hyperparameters on a held-out sequence of new tasks to confirm the numbers are not inflated by benchmark-specific tuning.
  • Because the readout is linear in random ReLU features, the method is essentially a kernel-style regressor; explicit kernel features such as random Fourier features might offer a principled way to tune the expansion dimension.
  • The framework could plausibly be combined with a small memory buffer for settings where class distributions shift more abruptly, though the paper does not test that combination.
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

4 major / 4 minor

Summary. The paper proposes an online incremental learning method for audio classification using a frozen PANNs CNN14 model. A random ReLU projection expands the 2048-dimensional embeddings to Q=8192, and ridge regression is applied to accumulated Gram and class-sum matrices (Eqs. 1-3). The method is evaluated on class-incremental learning with ESC-50 (93.4% final average accuracy) and domain-incremental learning with TAU Urban Acoustic Scenes 2019 (61.4%), and is compared to linear-probe, joint-linear-probe, and NCM baselines. The authors conclude the approach is suitable for both CIL and DIL and outperforms all investigated methods.

Significance. If the reported results hold, the paper offers a simple, base-task-free continual audio learner that leverages a strong pretrained representation and requires no iterative training. The ablation study in Figs. 3-4 provides useful evidence that the nonlinear random projection is the key component. The method is reproducible from the public PANNs model and standard datasets. However, the evaluation protocol and the absence of a direct comparison to RanPAC, the method on which it explicitly relies, currently prevent the broad superiority claim from being accepted at face value.

major comments (4)
  1. [§3.2, Fig. 4] The projection dimension Q=8192 is selected via ablation on exactly the same ESC-50 and TAU tasks that are later used to report the final accuracies. Because Fig. 4 shows average accuracy after each task (i.e., on the test splits), the headline numbers 93.4% and 61.4% are selected estimates, not unbiased evaluations of a pre-specified configuration. A held-out-task protocol or nested selection is needed; at minimum, the numbers should be labeled as having Q selected on the test tasks.
  2. [§2.2, Eq. (3)] The paper claims adaptation 'in a single forward pass' (Abstract and §2.2). However, the λ-selection procedure requires updating G and K for 17 candidate λ values on a random 80% subset of the current task, each involving the Q×Q inverse in Eq. (3), followed by a final update over the full task. This is not a single forward pass. If λ tuning is intended as separate hyperparameter selection, that should be stated explicitly and its computational cost quantified.
  3. [§2, §4, Table 1] The method is described as 'relies on RanPAC [24]', yet RanPAC is never run on the audio tasks, and no comparison to prior audio continual-learning methods (e.g., [4], [7], [10]) is provided. The conclusion that the method 'outperforms all other investigated methods' therefore only covers the simple baselines implemented in this paper. A direct RanPAC comparison under the same CIL/DIL protocols is necessary to support the claimed advantage.
  4. [Table 1] The advantage over the offline joint linear probe is 1.9 percentage points for CIL and 1.1 for DIL, with only n=5 runs. Given the small number of runs and the overlapping standard deviations (DIL: 61.4±0.1 vs 60.3±0.7), statistical significance testing or confidence intervals should be provided before making the 'outperforms' claim.
minor comments (4)
  1. [Abstract and §4] The wording 'outperforms all other investigated methods' should be qualified as 'baselines implemented in this paper' to avoid overclaiming.
  2. [Eq. (2)] K is defined as unnormalized class sums ('CP for class y without averaging'). In balanced datasets this is harmless, but for imbalanced tasks the lack of normalization can bias Wo. This design choice should be justified or at least noted.
  3. [§3.1] For the NCM baseline it is not stated whether cosine similarity is computed on the original 2048-dimensional embeddings or on the projected features. Please clarify, since Figs. 3-4 suggest the projection is central to the proposed method.
  4. [§3.2] The term 'online' is used to mean a single pass over the task data rather than a streaming/sample-by-sample setting; consider making this explicit in the definitions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the method is an empirical ridge-regression application with openly chosen hyperparameters.

full rationale

The paper's derivation chain is a direct application of ridge regression on random ReLU projections of frozen pretrained embeddings (Eqs. 1–3). The reported accuracies are evaluations, not predictions derived from the method's own outputs. Q is chosen from an ablation study on the same tasks, but this is transparent hyperparameter selection, not a fitted input renamed as a prediction; λ is selected on a validation split. The method explicitly relies on RanPAC [24], an external prior work, and self-citations concern related baselines rather than supporting the central claim. No step reduces an output to an input by definition or by self-citation.

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

The method rests on the discriminative power of frozen PANNs embeddings and on the RanPAC assumption that random ReLU projections make classes linearly separable. It introduces no new physical or mathematical entities. Two hyperparameters, Q and lambda, are fitted to data; W is random and frozen. The pretrained model and datasets are inputs from prior literature.

free parameters (2)
  • Projection dimension Q = 8192
    Selected by ablation on the same ESC-50 and TAU datasets (Fig. 4); increasing Q improves accuracy but raises computational cost.
  • Ridge regularization lambda = per task from grid 1e-8 to 1e8
    Selected per task by minimizing MSE on a 20% validation split of the current task (Section 2.2).
assumptions (4)
  • domain assumption PANNs CNN14 embeddings are discriminative and stable enough for classes and domains not seen in pretraining or in prior tasks.
    Core premise of the method; the frozen model is never adapted (Section 2.2).
  • domain assumption Random projection to higher dimension with ReLU preserves and enhances linear separability of classes.
    Relies on RanPAC [24] findings; not proven for audio embeddings (Eq. 1, Section 2.2).
  • standard math Ridge regression with unnormalized sums K and Gram matrix G provides a valid classifier, and lambda selected on validation transfers to test.
    Eq. (3) is standard ridge regression on one-hot targets; generalization of lambda selection is assumed.
  • standard math Random projection matrix W drawn from an unspecified distribution is adequate.
    No distribution or seed is given; the method assumes any typical random draw works.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online incremental learning for audio classification using a pretrained audio model." pith.science (2026). https://pith.science/paper/W463TRFX

@misc{pith2026250820732,
  author       = {Pith},
  title        = {Pith review of: Online incremental learning for audio classification using a pretrained audio model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W463TRFX}},
  note         = {Machine review of arXiv:2508.20732}
}
read the original abstract

Incremental learning aims to learn new tasks sequentially without forgetting the previously learned ones. Most of the existing incremental learning methods for audio focus on training the model from scratch on the initial task, and the same model is used to learn upcoming incremental tasks. The model is trained for several iterations to adapt to each new task, using some specific approaches to reduce the forgetting of old tasks. In this work, we propose a method for using generalizable audio embeddings produced by a pre-trained model to develop an online incremental learner that solves sequential audio classification tasks over time. Specifically, we inject a layer with a nonlinear activation function between the pre-trained model's audio embeddings and the classifier; this layer expands the dimensionality of the embeddings and effectively captures the distinct characteristics of sound classes. Our method adapts the model in a single forward pass (online) through the training samples of any task, with minimal forgetting of old tasks. We demonstrate the performance of the proposed method in two incremental learning setups: one class-incremental learning using ESC-50 and one domain-incremental learning of different cities from the TAU Urban Acoustic Scenes 2019 dataset; for both cases, the proposed approach outperforms other methods.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 27 canonical work pages

  1. [24]

    Ranpac: Random projections and pre-trained models for continual learning,

    M. D. McDonnell, D. Gong, A. Parvaneh, E. Abbasnejad, and A. Van den Hengel, “Ranpac: Random projections and pre-trained models for continual learning,” in NeurIPS, 2023

  2. [4]

    Class-incremental learning for multi- label audio classification,

    M. Mulimani and A. Mesaros, “Class-incremental learning for multi- label audio classification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 916–920

  3. [7]

    Online domain-incremental learning approach to classify acoustic scenes in all locations,

    M. Mulimani and A. Mesaros, “Online domain-incremental learning approach to classify acoustic scenes in all locations,” in European Signal Processing Conference (EUSIPCO). IEEE, 2024, pp. 96–100

  4. [10]

    Domain-incremental learning for audio classification,

    M. Mulimani and A. Mesaros, “Domain-incremental learning for audio classification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

  5. [1]

    Continual lifelong learning with neural networks: A review,

    G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, and S. Wermter, “Continual lifelong learning with neural networks: A review,” Neural networks, vol. 113, pp. 54–71, 2019

  6. [2]

    Adapter incremental continual learning of efficient audio spectrogram transformers,

    N. Muthuchamy Selvaraj, X. Guo, A. Kong, B. Shen, and A. Kot, “Adapter incremental continual learning of efficient audio spectrogram transformers,” in INTERSPEECH, 2023, pp. 909–913

  7. [3]

    Task incremental learning with static memory for audio classification without catastrophic interference,

    S. Karam, S.-J. Ruan, and Q. M. ul Haq, “Task incremental learning with static memory for audio classification without catastrophic interference,” IEEE Consumer Electronics Magazine, vol. 11, no. 5, pp. 101–108, 2022

  8. [5]

    Learning representations for new sound classes with continual self-supervised learning,

    Z. Wang, C. Subakan, X. Jiang, J. Wu, E. Tzinis, M. Ravanelli, and P. Smaragdis, “Learning representations for new sound classes with continual self-supervised learning,” IEEE Signal Processing Letters , vol. 29, pp. 2607–2611, 2022

Show all 28 references
  1. [6]

    Continual learning of new sound classes using generative replay,

    Z. Wang, C. Subakan, E. Tzinis, P. Smaragdis, and L. Charlin, “Continual learning of new sound classes using generative replay,” in 2019 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2019, pp. 308–312

  2. [8]

    Acoustic scene classification across multiple devices through incremental learning of device-specific domains,

    ——, “Acoustic scene classification across multiple devices through incremental learning of device-specific domains,” in Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE) , 2024, pp. 81–85

  3. [9]

    Three scenarios for continual learning,

    G. M. Van de Ven and A. S. Tolias, “Three scenarios for continual learning,” in Continual Learning Workshop NeurIPS , 2018

  4. [11]

    A closer look at class-incremental learning for multi-label audio classification,

    ——, “A closer look at class-incremental learning for multi-label audio classification,” IEEE Transactions on Audio, Speech and Language Processing, vol. 33, pp. 1293–1306, 2025

  5. [12]

    PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,

    Q. Kong, Y . Cao, T. Iqbal, Y . Wang, W. Wang, and M. D. Plumbley, “PANNs: Large-scale pretrained audio neural networks for audio pattern recognition,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 28, pp. 2880–2894, 2020

  6. [13]

    Ast: Audio spectrogram transformer,

    Y . Gong, Y .-A. Chung, and J. Glass, “Ast: Audio spectrogram transformer,” in INTERSPEECH, 2021, pp. 571–575

  7. [14]

    Efficient training of audio transformers with patchout,

    K. Koutini, J. Schl ¨uter, H. Eghbal-zadeh, and G. Widmer, “Efficient training of audio transformers with patchout,” in INTERSPEECH, 2022, pp. 2753–2757

  8. [15]

    SSAST: Self-supervised audio spectrogram transformer,

    Y . Gong, C.-I. Lai, Y .-A. Chung, and J. Glass, “SSAST: Self-supervised audio spectrogram transformer,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 36, no. 10, 2022, pp. 10 699–10 709

  9. [16]

    Enhancing acoustic scene classification with layer-wise fine-tuning on the SSAST model,

    S. Hao, D. Saito, and N. Minematsu, “Enhancing acoustic scene classification with layer-wise fine-tuning on the SSAST model,” in Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) . IEEE, 2024, pp. 1–6

  10. [17]

    CED: Consistent ensemble distillation for audio tagging,

    H. Dinkel, Y . Wang, Z. Yan, J. Zhang, and Y . Wang, “CED: Consistent ensemble distillation for audio tagging,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 291–295

  11. [18]

    Fine-tune the pretrained atst model for sound event detection,

    N. Shao, X. Li, and X. Li, “Fine-tune the pretrained atst model for sound event detection,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2024, pp. 911–915

  12. [19]

    Few- shot continual learning for audio classification,

    Y . Wang, N. J. Bryan, M. Cartwright, J. P. Bello, and J. Salamon, “Few- shot continual learning for audio classification,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021, pp. 321–325

  13. [20]

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

    Y . Li, W. Cao, W. Xie, J. Li, and E. Benetos, “Few-shot class-incremental audio classification using dynamically expanded classifier with self- attention modified prototypes,” IEEE Transactions on Multimedia, vol. 26, pp. 1346–1360, 2023

  14. [21]

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

    Y . Li, J. Li, Y . Si, J. Tan, and Q. He, “Few-shot class-incremental audio classification with adaptive mitigation of forgetting and overfitting,” IEEE/ACM Transactions on Audio, Speech, and Language Processing , vol. 32, pp. 2297–2311, 2024

  15. [22]

    Understanding intermediate layers using linear classifier probes,

    G. Alain and Y . Bengio, “Understanding intermediate layers using linear classifier probes,” arXiv preprint arXiv:1610.01644 , 2016

  16. [23]

    A simple episodic linear probe improves visual recognition in the wild,

    Y . Liang, L. Zhu, X. Wang, and Y . Yang, “A simple episodic linear probe improves visual recognition in the wild,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 2022, pp. 9559–9569

  17. [25]

    ESC: Dataset for Environmental Sound Classification,

    K. J. Piczak, “ESC: Dataset for Environmental Sound Classification,” in Proceedings of the 23rd Annual ACM Conference on Multimedia , 2015, pp. 1015–1018

  18. [26]

    Environmental sound classification with convolutional neural networks,

    ——, “Environmental sound classification with convolutional neural networks,” in 25th IEEE International Workshop on Machine Learning for Signal Processing (MLSP) . IEEE, 2015, pp. 1–6

  19. [27]

    A multi-device dataset for urban acoustic scene classification,

    A. Mesaros, T. Heittola, and T. Virtanen, “A multi-device dataset for urban acoustic scene classification,” in Workshop on Detection and Classification of Acoustic Scenes and Events (DCASE) , 2018, pp. 9–13

  20. [28]

    SGDR: Stochastic gradient descent with warm restarts,

    I. Loshchilov and F. Hutter, “SGDR: Stochastic gradient descent with warm restarts,” in International Conference on Learning Representations (ICLR), 2017

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.