Pith. sign in

REVIEW 3 major objections 5 minor 26 references

Class-Incremental Learning for Sound Event Localization and Detection

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read An MSE distillation loss lets a sound-event localization and detection model add four new sound classes while keeping baseline-level detection and improving localization.

desk verdict First CIL-SELD paper with a real effect, but the headline comparison to 'baseline' is not controlled and needs a matched baseline before the central claim holds. read the letter →

arxiv 2411.12830 v1 pith:PPJHZZ5H submitted 2024-11-19 eess.AS

Ruchi Pandey , Manjunath Mulimani , Archontis Politis , Annamaria Mesaros This is my paper · ORCID
classification eess.AS
keywords class-incrementallearningsoundeventlocalizationanddetectioncatastrophicforgettingoutputdistillationACCDOATAU-NIGENScontinualpolyphonicaudio
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

This paper asks whether a sound event localization and detection (SELD) model can be taught four new sound classes after deployment without forgetting the eight classes it already knows, and without retraining on old-class recordings. The proposed two-stage method, CIL-MSE, keeps the first-stage model's outputs for the original classes as a distillation target while the expanded model learns the new classes from new data. On the TAU-NIGENS Spatial Sound Events 2021 dataset, the method matches the reference baseline's overall F1-score and error rate, and even improves localization error. The result matters because real-world audio systems are often deployed with a fixed class set and later need to grow; if this holds, such growth can be done cheaply without catastrophic forgetting.

What carries the argument

The load-bearing mechanism is the MSE-based output distillation loss, adapted from an earlier class-incremental learning method for multi-label audio classification [19]. At stage 1, the frozen stage-0 model's predictions for the original 8 classes serve as soft targets; the active model is trained so that its outputs for those classes remain close to the frozen outputs, while a standard MSE loss supervises the 4 new classes. The two terms are combined as $L = (1-\lambda) L_{\mathrm{MSE}} + \lambda L_{\mathrm{OD}}$, with $\lambda = 0.5$ in the main result. Because the regression head uses ACCDOA vectors, the distillation penalty acts directly on the joint detection-and-localization output space rather than on class probabilities.

What would settle it

Train the same architecture that the incremental model uses (the DCASE 2021 CRNN) jointly on all 12 classes with the same hyperparameters and evaluate it on the same test split; if that model clearly outperforms CIL-MSE in F1 or error rate, the paper's claim of matching baseline performance would no longer stand.

Watch

Extended reading notes

Core claim

The paper's central claim is that output-distillation-based class-incremental learning can add four new sound classes to an eight-class SELD model while preserving the original classes' detection and localization performance. Using the ACCDOA representation, in which each active sound event is encoded as a three-dimensional Cartesian direction-of-arrival vector, the stage-1 model expands the final regression layer from $8\times3$ to $12\times3$ neurons. It is trained with a total loss that blends the mean-square-error loss for the new classes with a mean-square-error distillation loss on the old classes' outputs, $L = (1-\lambda)L_{\mathrm{MSE}} + \lambda L_{\mathrm{OD}}$. With the mixing weight $\lambda=0.5$, the resulting CIL-MSE model achieves overall F1 of 23.1, error rate 0.71, localization error 26.6°, and localization recall 40.9%, compared to the reference baseline's 23.1, 0.70, 39.4°, and 41.2%. The authors conclude that CIL-SELD performs similarly to the baseline, showing improved localization for both existing and newly introduced classes without training on the entire dataset.

Load-bearing premise

The claim that CIL-MSE maintains baseline performance rests on comparing against a DCASE 2022 baseline built on a different architecture than the DCASE 2021 backbone of the incremental model, with no same-architecture 12-class model trained as a matched control.

Editorial extensions

If this is right

  • A deployed SELD system can absorb new sound classes in an incremental stage without access to labelled recordings of the old classes, cutting the data and compute cost of model updates.
  • Because the classifier head expands by a fixed block of neurons, the same two-stage recipe extends to multiple incremental steps, which the paper names as future work.
  • MSE distillation beats KL-divergence distillation on overall F1 in this setup, so the choice of distillation loss is a deciding factor in whether old-class knowledge survives.
  • The balanced L2 norms of the final-layer regression weights under CIL-MSE, in contrast to fine-tuning's skewed norms, indicate the model distributes capacity evenly across old and new classes.

Reading between the lines

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

  • The comparison baseline is the DCASE 2022 SELD baseline, while the incremental model uses the DCASE 2021 backbone; a same-architecture jointly-trained 12-class model would make the 'maintains baseline' claim directly testable.
  • The observed localization-error improvement (26.6° vs 39.4°) may stem partly from the architecture difference rather than from incremental learning; a matched baseline would isolate the effect of distillation.
  • If the method is extended to many small incremental steps, distillation targets would be refreshed at each stage, but the accumulation of distillation error over multiple stages is an open question the paper does not address.
  • Because the distillation loss operates on continuous regression outputs, the same idea could transfer to other multi-output audio regression tasks such as direction-of-arrival estimation alone.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a two-stage class-incremental learning (CIL) framework for sound event localization and detection (SELD). In Stage 0, a CRNN based on the DCASE 2021 SELD baseline is trained on 8 sound classes using the ACCDOA representation. In Stage 1, the classifier head is expanded to 12 classes and the model is trained on data for the 4 new classes while an output distillation loss (MSE) penalizes discrepancies between the old and updated models' outputs for the original 8 classes. Experiments on TAU-NIGENS Spatial Sound Events 2021 report that the proposed CIL-MSE (lambda=0.5) approximately matches a DCASE 2022 Challenge SELD baseline on overall F1, ER, LE, and LR, and outperforms fine-tuning and independent learning, with additional ablations on distillation loss type and lambda.

Significance. If the results are substantiated, this is the first demonstration of class-incremental learning for the joint detection and localization of sound events, which is a meaningful step for adaptive SELD systems that need to add classes without full retraining. The paper also includes useful ablations (MSE vs KL distillation, lambda sweep, class-wise analysis, and analysis of regression weight norms) that go beyond a single accuracy number. However, the central claim that CIL 'maintains baseline performance' is currently supported only by an unmatched baseline comparison, and the choice of lambda=0.5 is not adequately justified. These issues are load-bearing and must be addressed before the claim can be accepted.

major comments (3)
  1. [Section II vs Section III.C, Table I] The comparison that supports the central claim is not controlled. The proposed CIL model is built on the DCASE 2021 baseline CRNN architecture [24], while the 'Baseline' in Table I is the DCASE 2022 Challenge SELD baseline [25]. These are different architectures, feature pipelines, and training configurations. The paper never trains the same architecture used for CIL on all 12 classes jointly and reports its performance. Therefore the observation that CIL-MSE is close to the DCASE 2022 baseline (and even improves LE by 12.8 degrees) could be due to architectural differences rather than the incremental method. To support the claim that CIL 'maintains baseline performance', the authors must supply a matched baseline: the same CRNN/feature extractor trained jointly on all 12 classes, with the same optimizer, epochs, and evaluation protocol.
  2. [Section IV.B, Figure 3] The choice of lambda=0.5 is not justified with respect to data leakage. Figure 3 shows average F1 over all 12 classes as a function of lambda, and the paper selects lambda=0.5 for the reported CIL-MSE results. It is not stated whether this sweep was performed on the validation split or on the same test split used in Tables I and II. If the sweep and the final evaluation use the same test partition, the reported numbers are the result of tuning on the test set, which overstates the method's expected performance. Please clarify the data split used for Figure 3 and, preferably, report results for a fixed lambda chosen on validation only. In addition, no error bars or multiple-seed runs are reported; given the small metric differences involved, this is needed to establish that the reported ordering is not noise.
  3. [Table II] The paper's claim that CIL preserves old-class knowledge cannot be evaluated from the presented tables. Table II reports old-class (1-8) and new-class (9-12) F1 for FT, IndL, CIL-MSE, and CIL-KLD, but it does not include the corresponding old-class and new-class breakdown for the baseline model. Without the baseline's old-class F1, the reader cannot tell whether CIL-MSE's old-class F1 of 19.7 represents retention comparable to joint training or a substantial drop. Please add the same-architecture baseline's old-class and new-class F1 scores (and ideally the per-class values) to Table II.
minor comments (5)
  1. [Section III.C] Reference [25] is cited as the source of the DCASE 2022 SELD baseline, but [25] is the STARSS22 dataset paper; the actual DCASE 2022 baseline system should be cited explicitly, or the reference should be corrected.
  2. [Equation (1) and Figure 1] The distillation loss is denoted inconsistently: the text and equation use LOD with a subscript, while the figure uses what appears to be a calligraphic L_OD. Please unify the notation throughout.
  3. [Section IV.A] The statement that LE bars are omitted for classes with zero F1 'as LE for these cases is 180 degrees' is confusing, since localization error is normally undefined or set to a penalty value when no true positive exists. Please specify the exact convention used in the evaluation code.
  4. [Section IV, paragraph after Table I] The phrase 'probably due to its MSE-based distillation loss which directly minimizes discrepancies in logits' is speculative; either remove 'probably' or provide supporting evidence (e.g., an analysis of output distributions or per-class errors).
  5. [Section I heading] The heading 'I. I N TRODUCTION' contains an extra space in 'INTRODUCTION'; this is a minor typographical issue.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the main claim is an empirical comparison against an external baseline and a held-out test set; self-citations to prior distillation work are method references, not load-bearing.

full rationale

The paper's central result is an empirical observation: after incremental training with Eq. (1), the model's SELD metrics on the TAU-NIGENS 2021 test partition are comparable to the DCASE 2022 baseline (Table I). Nothing in the derivation makes this result true by construction. The supervised term LMSE uses ground-truth ACCDOA targets for the four new classes; the distillation term LOD compares the current model's outputs for the old eight classes to the frozen Stage-0 model's outputs. Neither term is defined in terms of the reported ER/F1/LE/LR metrics, and the evaluation metrics are computed against held-out test annotations. The closest thing to a circularity concern is that the output-distillation loss is 'similar to [19]' and the IndL comparison is from [20], both by the same research group. This is a method reference, not a load-bearing self-citation: no uniqueness theorem, no ansatz-via-citation, and no external result is replaced by the authors' prior claim. The baseline is an external DCASE 2022 baseline [25]; the fact that the CIL model uses a DCASE 2021-style CRNN is a potential confound for the comparison, but that is a fairness/validity issue, not circular reasoning. The λ=0.5 value shown in Table I is selected after the sweep in Fig. 3, which introduces mild selection pressure, but selecting a hyperparameter by validation/test performance is not equivalent to fitting then renaming a prediction. No equation in the paper reduces to its inputs, so no circular step can be exhibited.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central result rests on standard SELD representations and metrics, plus an ad-hoc assumption that the DCASE 2022 baseline is comparable to the DCASE 2021-based model. The only free parameter is the distillation weight lambda, and its selection source is not fully specified.

free parameters (1)
  • lambda (λ) = 0.5 (main result), range 0.1-0.9 explored
    Trade-off weight between new-class MSE loss and old-class distillation loss in Eq. (1). The main results use lambda=0.5, chosen based on the curve in Figure 3; the paper does not state whether the curve is computed on validation or test data.
assumptions (4)
  • domain assumption The ACCDOA representation, where each active event is encoded as a 3D Cartesian DOA vector and activity is thresholded on vector magnitude, is a valid formulation for SELD.
    The entire method and evaluation depend on this representation from [6]. The paper does not re-derive it but treats it as given.
  • domain assumption The two-stage curriculum (8 classes then 4 new classes) is representative of real deployment needs and does not introduce bias into the evaluation.
    The paper selects this split without justifying it beyond saying it is efficient. The choice affects the difficulty of the incremental step and the interpretation of the results.
  • ad hoc to paper The DCASE 2022 baseline model [25] is a comparable reference to the proposed CIL model built on the DCASE 2021 baseline [24].
    Section III.C introduces this baseline without acknowledging the architecture mismatch with Section II. This assumption is load-bearing for the conclusion that CIL 'maintains baseline performance'.
  • domain assumption The evaluation metrics (ER, F1, LE, LR) with threshold T=20 degrees correctly capture SELD performance.
    The paper uses these metrics from [7] without discussion. The threshold choice is standard in the field.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Class-Incremental Learning for Sound Event Localization and Detection." pith.science (2026). https://pith.science/paper/PPJHZZ5H

@misc{pith2026241112830,
  author       = {Pith},
  title        = {Pith review of: Class-Incremental Learning for Sound Event Localization and Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPJHZZ5H}},
  note         = {Machine review of arXiv:2411.12830}
}
read the original abstract

This paper investigates the feasibility of class-incremental learning (CIL) for Sound Event Localization and Detection (SELD) tasks. The method features an incremental learner that can learn new sound classes independently while preserving knowledge of old classes. The continual learning is achieved through a mean square error-based distillation loss to minimize output discrepancies between subsequent learners. The experiments are conducted on the TAU-NIGENS Spatial Sound Events 2021 dataset, which includes 12 different sound classes and demonstrate the efficacy of proposed method. We begin by learning 8 classes and introduce the 4 new classes at next stage. After the incremental phase, the system is evaluated on the full set of learned classes. Results show that, for this realistic dataset, our proposed method successfully maintains baseline performance across all metrics.

Figures

Figures reproduced from arXiv: 2411.12830 by the authors.

Figure 1
Figure 1. Two stage class-incremental learning for SELD. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Class-wise F1-scores and localization error for different methods. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. L2-norm of the regression weight vectors [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

  1. [25]

    Starss22: A dataset of spatial recordings of real scenes with spatiotem- poral annotations of sound events,

    A. Politis, K. Shimada, P. Sudarsanam, S. Adavanne, D. Krause, Y . Koyama, N. Takahashi, S. Takahashi, Y . Mitsufuji, and T. Virtanen, “Starss22: A dataset of spatial recordings of real scenes with spatiotem- poral annotations of sound events,” arXiv preprint arXiv:2206.01948 , 2022

    arXiv 2022

  2. [24]

    A dataset of dynamic reverberant sound scenes with directional interferers for sound event localization and detection,

    A. Politis, S. Adavanne, D. Krause, A. Deleforge, P. Srivastava, and T. Virtanen, “A dataset of dynamic reverberant sound scenes with directional interferers for sound event localization and detection,” arXiv preprint arXiv:2106.06999, 2021

    arXiv 2021

  3. [19]

    Class-incremental learning for multi- label audio classification,

    M. Mulimani and A. Mesaros, “Class-incremental learning for multi- label audio classification,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces., 2024, pp. 916–920

    work page 2024

  4. [20]

    Incremental Learning of Acoustic Scenes and Sound Events

    ——, “Incremental learning of acoustic scenes and sound events,” arXiv preprint arXiv:2302.14815, 2023

    work page Pith review arXiv 2023

  5. [1]

    Direction of arrival estima- tion for multiple sound sources using convolutional recurrent neural network,

    S. Adavanne, A. Politis, and T. Virtanen, “Direction of arrival estima- tion for multiple sound sources using convolutional recurrent neural network,” in Euro. Sig. Proces. Conf. (EUSIPCO), 2018, pp. 1462–1466

    work page 2018

  6. [2]

    Sound event localization and detection of overlapping sources using convolutional recurrent neural networks,

    S. Adavanne, A. Politis, J. Nikunen, and T. Virtanen, “Sound event localization and detection of overlapping sources using convolutional recurrent neural networks,” IEEE J. Sel. Topics Signal Process. , vol. 13, no. 1, pp. 34–48, 2018

    work page 2018

  7. [3]

    The application of DOA estimation approach in patient tracking systems with high patient density,

    L. Wan, G. Han, L. Shu, S. Chan, and T. Zhu, “The application of DOA estimation approach in patient tracking systems with high patient density,” IEEE Trans. Ind. Electron. , vol. 12, no. 6, pp. 2353–2364, 2016

    work page 2016

  8. [4]

    Maximum likelihood approach to “informed

    M. Farmani, M. S. Pedersen, Z. H. Tan, and J. Jensen, “Maximum likelihood approach to “informed” sound source localization for hearing aid applications,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces. , 2015, pp. 16–20

    work page 2015

Show all 26 references
  1. [5]

    A survey of sound source localization with deep learning methods,

    P. A. Grumiaux, S. Kiti ´c, L. Girin, and A. Gu ´erin, “A survey of sound source localization with deep learning methods,” J. Acoust. Soc. Am. , vol. 152, no. 1, pp. 107–151, 2022

    2022

  2. [6]

    Accdoa: Activity-coupled cartesian direction of arrival representation for sound event localization and detection,

    K. Shimada, Y . Koyama, N. Takahashi, S. Takahashi, and Y . Mitsufuji, “Accdoa: Activity-coupled cartesian direction of arrival representation for sound event localization and detection,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces. , 2021, pp. 915–919

    2021

  3. [7]

    Joint measurement of localization and detection of sound events,

    A. Mesaros, S. Adavanne, A. Politis, T. Heittola, and T. Virtanen, “Joint measurement of localization and detection of sound events,” in IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), 2019, pp. 333–337

    2019

  4. [8]

    An improved event-independent network for polyphonic sound event localization and detection,

    Y . Cao, T. Iqbal, Q. Kong, F. An, W. Wang, and D. M. Plumbley, “An improved event-independent network for polyphonic sound event localization and detection,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces., 2021, pp. 885–889

    2021

  5. [9]

    Polyphonic sound event detection using convolutional bidirectional lstm and synthetic data-based transfer learning,

    S. Jung, J. Park, and S. Lee, “Polyphonic sound event detection using convolutional bidirectional lstm and synthetic data-based transfer learning,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces., 2019, pp. 885– 889

    2019

  6. [10]

    Semi-supervised sound event detection with pre-trained model,

    L. Xu, L. Wang, S. Bi, H. Liu, and J. Wang, “Semi-supervised sound event detection with pre-trained model,” in IEEE Inter . Conf. Acous., Spe., Sig. Proces. , 2023, pp. 1–5

    2023

  7. [11]

    Dual knowledge distillation for efficient sound event detection,

    Y . Xiao and R. K. Das, “Dual knowledge distillation for efficient sound event detection,” arXiv preprint arXiv:2402.02781 , 2024

    2024 arXiv

  8. [12]

    Learning a unified clas- sifier incrementally via rebalancing,

    S. Hou, X. Pan, C. Loy, Z. Wang, and D. Lin, “Learning a unified clas- sifier incrementally via rebalancing,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2019, pp. 831– 839

    2019

  9. [13]

    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

    2019

  10. [14]

    icarl: Incremental classifier and representation learning,

    S. A. Rebuffi, A. Kolesnikov, G. Sperl, and C. H. Lampert, “icarl: Incremental classifier and representation learning,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2017, pp. 2001–2010

    2017

  11. [15]

    Three scenarios for continual learning,

    G. M. V . de Ven and A. S. Tolias, “Three scenarios for continual learning,” arXiv preprint arXiv:1904.07734 , 2019

    1904 arXiv

  12. [16]

    Learning without forgetting,

    Z. Li and D. Hoiem, “Learning without forgetting,” IEEE transactions on pattern analysis and machine intelligence , vol. 40, no. 12, pp. 2935– 2947, 2017

    2017

  13. [17]

    Multi-domain incremental learning for semantic segmen- tation,

    P. Garg, R. Saluja, V . Balasubramanian, C. Arora, A. Subramanian, and C. Jawahar, “Multi-domain incremental learning for semantic segmen- tation,” in IEEE/CVF Winter Conference on Applications of Computer Vision, 2022, pp. 761–771

    2022

  14. [18]

    Charac- terizing continual learning scenarios and strategies for audio analysis,

    R. Bhatt, P. Kumari, D. Mahapatra, A. E. Saddik, and M.Saini, “Charac- terizing continual learning scenarios and strategies for audio analysis,” arXiv preprint arXiv:2407.00465 , 2024

    2024 arXiv

  15. [21]

    Progressive continual learning for spoken keyword spotting,

    Y . Huang, N. Hou, and N. F. Chen, “Progressive continual learning for spoken keyword spotting,” in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2022, pp. 7552–7556

    2022

  16. [22]

    Continual learning for on-device environmental sound classification,

    Y . Xiao, X. Liu, J. King, A. Singh, E. S. Chng, M. D. Plumbley, and W. Wang, “Continual learning for on-device environmental sound classification,” arXiv preprint arXiv:2207.07429 , 2022

    2022 arXiv

  17. [23]

    Configurable doa estimation using incremental learning,

    Y . Xiao and R. K. Das, “Configurable doa estimation using incremental learning,” arXiv preprint arXiv:2407.03661 , 2024

    2024 arXiv

  18. [26]

    Overview and evaluation of sound event localization and detection in dcase 2019,

    A. Politis, A. Mesaros, S. Adavanne, T. Heittola, and T. Virtanen, “Overview and evaluation of sound event localization and detection in dcase 2019,” IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 29, pp. 684–698, 2020

    2019

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.