Pith. sign in

REVIEW 2 major objections 7 minor 39 references

Greedy Volume Maximization of Gradient Embeddings for Long-Tailed Frame-Level Bioacoustic Active Learning

T0 review · 2 major / 7 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read The paper claims that greedily maximizing the volume spanned by residual-weighted BADGE gradient embeddings yields the best active-learning performance on a long-tailed, frame-level hyena call-type task, with its largest margins on the rare

desk verdict A clean, well-tested combination of known machinery for long-tailed bioacoustic active learning; the greedy traversal clearly earns its keep, but the frame-wise embedding contribution is never isolated, so the paper needs one ablation before I'd call it complete. read the letter →

arxiv 2607.13555 v1 pith:IWME3UZC submitted 2026-07-15 eess.AS cs.AIcs.SD

Shiqi Zhang , Marius Fai{ss} , Ariana Strandburg-Peshkin , Tuomas Virtanen This is my paper · ORCID
classification eess.AScs.AIcs.SD
keywords activelearningframe-levelbioacousticclassificationBADGEdeterminantalpointprocessgreedysubmodularmaximizationlong-tailedclassdistributionrarecalltypeslabelefficiency
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 proposes BADGE-Greedy-DPP, an active-learning query strategy for frame-level bioacoustic call-type classification where target calls are extremely sparse and the class distribution is long-tailed. It replaces the sampling-based batch construction in BADGE with a deterministic greedy rule that adds the segment whose gradient embedding most enlarges the volume spanned by the batch, an objective that is monotone submodular and therefore carries a (1−1/e) approximation guarantee. The paper also builds segment embeddings frame-wise, weighting each frame's pseudo-gradient by its prediction residual so that a few uncertain rare-call frames can set a segment's direction instead of being washed out by abundant confidently predicted no-call frames. On a hyena dataset across 10 runs, it reports the best overall and rare-call-type performance among all compared strategies, and is the only method to reach the full-supervised reference in every run, doing so after about 3.94% of the training set.

What carries the argument

The central object is the regularized log-determinant volume objective F(S) = log det(λI + Φ_S^T Φ_S), where Φ_S rows are the gradient embeddings of the selected segments. The greedy rule adds the candidate with the largest marginal gain Δ(i|S) = log(1 + φ_i^T (λI + Φ_S^T Φ_S)^{-1} φ_i), which measures how much new gradient direction the candidate contributes beyond the current batch span. Because F is monotone submodular, greedy selection guarantees at least a (1−1/e) fraction of the optimum of this proxy objective. The embedding is constructed frame-wise as G_i = (1/T) Σ_t (p_it − ˆy_it) ⊗ h_it, where each frame contributes a pseudo-gradient weighted by its prediction residual, so uncertai

What would settle it

Run BADGE-Greedy-DPP with the residual-weighted embedding of Eq. (2) versus the same greedy traversal using uniformly averaged frame features or unweighted frame-wise gradients on the same hyena pool; if the Rare-N-AULC gap collapses or reverses, the claimed granularity-mismatch mechanism is not the cause of the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that greedy maximization of the regularized log-determinant of BADGE gradient embeddings, together with frame-wise residual-weighted aggregation, resolves the temporal granularity mismatch in frame-level active learning and yields the strongest label efficiency on a long-tailed hyena call-type task. The authors would state it as: BADGE-Greedy-DPP achieves higher N-AULC, final mAP, and rare-call-type metrics than MFFT and both vanilla BADGE traversals, with the largest margins on the rarest classes, while being deterministic, computationally cheaper than MCMC-based BADGE, and carrying a worst-case approximation guarantee that the k-means++ and MCMC DPP heuristics do not p

Load-bearing premise

The claim rests on the assumption that residual-weighted frame-wise aggregation preserves and amplifies the signal of rare uncertain frames better than uniform averaging; if uniform averaging performed equally well under the same greedy traversal, the paper's proposed mechanism and part of its rationale for the rare-class gains would be unsupported.

Editorial extensions

If this is right

  • If the central claim is correct, a label budget of 3000 segments suffices for the greedy selector to match full-supervised performance in all 10 runs, at a median of 2040 labeled segments (3.94% of the training set), whereas the strongest baseline reaches it in 90% of runs with 2533 segments.
  • Rare call types benefit the most: the reported Rare-N-AULC is 47.4% versus 43.6% for MFFT, and enrichment is more balanced across the three rarest call types rather than concentrated on one or two.
  • The (1−1/e) approximation guarantee distinguishes the method from BADGE's k-means++ and MCMC DPP traversals, making batch selection deterministic and reproducible while remaining fast: 142 seconds per full active-learning run versus 769 seconds for MCMC DPP.
  • Because BADGE already encodes uncertainty through pseudo-gradients, the greedy volume rule needs no separate weight balancing uncertainty and diversity, simplifying the acquisition function.
  • If the approach generalizes, any frame-level multi-label sparse detection task could adopt the same residual-weighted embedding and greedy volume traversal without changing the classifier.

Reading between the lines

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

  • The paper does not isolate the frame-wise residual weighting from the traversal rule: all BADGE variants share the same embedding, so the reported gains could be driven entirely by greedy volume selection; a uniform-average embedding under the same greedy rule would settle which contribution matters.
  • The submodularity guarantee is a guarantee about the proxy volume objective, not about test mAP; on datasets where the gradient embeddings are less informative, the empirical margin could shrink even though the guarantee remains intact.
  • The same greedy DPP traversal could be applied to other sparse, long-tailed audio event detection tasks with segment-level annotation, and the 'single uncertain rare frame sets direction' mechanism suggests a further extension: explicitly rewarding segments predicted to contain rare-class frames.
  • The paper's framing implies that active learning for bioacoustics should be evaluated on rare-class metrics and full-supervision reach rates, not just final mAP; that view, if adopted more broadly, would change how query strategies are benchmarked.
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

2 major / 7 minor

Summary. The paper proposes BADGE-Greedy-DPP, a batch active-learning selector for long-tailed frame-level bioacoustic classification. It represents each 10-s segment by a frame-wise residual-weighted BADGE gradient embedding (Eq. (2)) and selects batches by greedily maximizing the regularized log-determinant (volume) of selected embeddings (Eq. (3)). Since F(S) is monotone submodular, the greedy rule gives a (1−1/e) guarantee for this proxy objective, which the k-means++ and MCMC DPP BADGE traversals lack. The method is evaluated on a hyena call-type dataset over 10 runs; it reports best N-AULC, final mAP, rare-class mAP, and enrichment among the compared strategies, with Holm-corrected permutation tests.

Significance. The theoretical part is sound: log-determinant objectives of the form (3) are standard submodular functions and the greedy approximation guarantee is correctly cited. The empirical protocol is careful—10 independent runs, min-max bands, corrected permutation tests, and a full-supervised reference—and the proposed method does show consistent gains on the studied dataset. If the embedding mechanism were isolated, the paper would be a solid contribution to bioacoustic active learning. However, the second contribution (frame-wise residual weighting) is not tested independently of the greedy traversal, and the empirical evidence is confined to a single dataset. No code or data are provided.

major comments (2)
  1. [Sec. 3.1, Eq. (2), Table 2] Section 3.1 states that all BADGE variants use the same pseudo-gradient embedding construction (2) and differ only in the traversal step. Consequently, the comparisons in Table 2 isolate the greedy traversal but provide no evidence for the second claimed contribution—the frame-wise residual weighting. The abstract and Section 1 assert that this weighting 'lets a single uncertain rare-call frame set the segment's direction' and that uniform averaging 'washes them out.' No experiment replaces Eq. (2) with a uniform frame average or a segment-level pseudo-gradient while keeping the greedy traversal. This is load-bearing: without such an ablation, the rare-class gains (e.g., snr enrichment in Fig. 2) cannot be attributed to the residual-weighting mechanism. Please add an ablation comparing Greedy-DPP with (i) Eq. (2), (ii) uniformly averaged frame features, and (iii) a segment-level pseudo-g
  2. [Sec. 4, Table 2] The empirical evaluation is on a single task: a preliminary version of HyenaSET. All headline claims ('BADGE-Greedy-DPP achieves the best overall and rare-call-type performance') are conditional on this one dataset, which is co-authored by a member of the research team. The abstract and title generalize to long-tailed frame-level bioacoustic active learning. Please add at least one independent dataset (or a well-motivated synthetic benchmark) to demonstrate that the gains are not dataset-specific. If this is not feasible, the claims should be explicitly reframed as a case study.
minor comments (7)
  1. [Title page] 'V olume' should be 'Volume'; 'Under reviewing' should be 'Under review'.
  2. [Eq. (6)] 'N-AU LC' has a spacing typo; should be 'N-AULC'.
  3. [Sec. 4.1] The text reports a final count of selected segments carrying rare-call labels (496.8±11.3), but this value is not shown in Table 2. Either add it to the table or remove the unreferenced number.
  4. [Sec. 2.1/3.1] The sentence 'We change only the traversal rule applied to these embeddings' is ambiguous because the paper also proposes the embedding itself. Clarify that this statement refers to the comparison among BADGE variants.
  5. [Sec. 2.2, Eq. (3)] The approximation guarantee is correctly stated as being for the proxy objective F(S). However, the introduction's phrasing 'guarantee not provided by BADGE's existing heuristics' could mislead readers into thinking the guarantee is about final classifier performance. Please state explicitly that the greedy bound does not directly bound test mAP.
  6. [Sec. 3.2/Fig. 2] Fig. 2 plots per-class enrichment, while R-Enr in Table 2 is a macro average; the caption should clarify this to avoid confusion.
  7. [Sec. 4.2] FS-Bud/FS-Rch compare active-learning runs against the mean full-supervised mAP, ignoring the reference variance (62.2±1.5). This is acceptable, but it could be more robust; mentioning this as a limitation would be appropriate.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: submodularity guarantee is external and empirical gains are measured; missing ablation is an experimental gap, not circularity.

full rationale

The paper's derivation chain is not circular. The proposed selector's approximation guarantee (Section 2.2) rests on standard external results: monotone submodularity of the regularized log-determinant and the (1-1/e) greedy bound [31]-[33],[38], none of which are authored by this team. The BADGE gradient embedding in Eq. (2) follows the original BADGE [28]; the frame-wise residual weighting is a design choice whose stated property (confident frames contribute little) is true by construction, but the paper does not present that property as a fitted prediction. The central empirical claim - best N-AULC, rare-N-AULC, F-mAP, and F-rmAP - is measured on a held-out test split against external baselines (Random, Entropy, Farthest Traversal, Disagreement, MFFT) and two vanilla BADGE traversals; the gains are not derived from the equations themselves. Self-citations appear only as baselines (MFFT [36],[37]) or as dataset/embedding resources (HyenaSET [34], animal2vec [35]) and are not used to justify the method's correctness. The lack of an ablation isolating the frame-wise residual weighting (all BADGE variants share Eq. (2)) is a genuine experimental gap - the rare-class gains could in principle be due to the greedy traversal alone - but that is a question of attribution/evidence, not circular reasoning: no fitted parameter is renamed as a prediction, and no conclusion reduces to its input by construction.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. It uses two fixed hyperparameters (λ and the 0.5 pseudo-label threshold) and several domain assumptions inherited from BADGE and from the authors' dataset. The main unverified premise is the frame-wise residual weighting mechanism, which is not ablated.

free parameters (2)
  • λ (log-det regularizer) = 1e-6
    Small ridge term for invertibility and numerical stability of the log-determinant objective (Eq. 3/5); fixed, not tuned to data.
  • pseudo-label threshold = 0.5
    Threshold in Eq. (1) to form hallucinated labels from posterior; standard default, not fitted.
assumptions (5)
  • standard math F(S) = log det(λI + Φ_S^T Φ_S) is monotone submodular, so greedy achieves (1−1/e) of the optimum for maximizing F
    Invoked in Section 2.2, citing [31–33,38]; standard result.
  • standard math Matrix determinant lemma gives the marginal gain formula (5)
    Used in Eq. (5).
  • domain assumption Pseudo-labels from thresholding predictions at 0.5 produce useful gradient embeddings for uncertainty/diversity (BADGE assumption)
    Borrowed from BADGE [28]; it is an assumption that these embeddings capture informative directions.
  • domain assumption Frame-wise residual weighting (Eq. 2) yields a segment embedding whose direction is set by uncertain frames; uniform averaging would wash them out
    Core intuition in Section 2.1; not formally proven and not isolated by ablation.
  • domain assumption The HyenaSET preliminary dataset and animal2vec embeddings are representative and suitable for evaluating active learning
    Used for all experiments; dataset from authors' collaboration.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Greedy Volume Maximization of Gradient Embeddings for Long-Tailed Frame-Level Bioacoustic Active Learning." pith.science (2026). https://pith.science/paper/IWME3UZC

@misc{pith2026260713555,
  author       = {Pith},
  title        = {Pith review of: Greedy Volume Maximization of Gradient Embeddings for Long-Tailed Frame-Level Bioacoustic Active Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IWME3UZC}},
  note         = {Machine review of arXiv:2607.13555}
}
read the original abstract

Bioacoustic call-type classification relies on costly expert annotation. Active learning can reduce this burden by selecting a small batch of segments for expert annotation and using the labeled segments for training the classifier. The setting is hard: the target calls are extremely sparse and the call-type distribution is long-tailed, so a tight budget must be spent on the few rare, informative segments. We propose BADGE-Greedy-DPP, a deterministic batch selector that greedily adds the segment whose BADGE gradient embedding most enlarges the volume spanned by the batch; because this log-volume objective is submodular, the greedy rule guarantees a batch value at least a (1-1/e) fraction of the optimum of this objective, a guarantee not provided by BADGE's existing k-means++ and MCMC DPP sampling heuristics. There is also a temporal granularity mismatch in the task. The acquisition function scores whole segments, yet the informative frames inside them are few. Uniform averaging therefore washes them out. We show that the BADGE construction naturally addresses this mismatch when applied frame-wise, as prediction residuals weight the aggregated pseudo-gradient, so confidently predicted no-call frames contribute little while a single uncertain rare-call frame can still set the segment's direction. Across 10 runs on a sparse, imbalanced hyena call-type dataset, BADGE-Greedy-DPP achieves the best overall and rare-call-type performance among all compared query strategies, including MFFT, the strongest non-BADGE baseline, and the two vanilla BADGE traversals.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 2 linked inside Pith

  1. [1]

    Computational bioacoustics with deep learning: a review and roadmap,

    D. Stowell, “Computational bioacoustics with deep learning: a review and roadmap,”PeerJ, vol. 10, p. e13152, 2022

    2022

  2. [2]

    Active learning for sound event detection,

    Z. Shuyang, T. Heittola, and T. Virtanen, “Active learning for sound event detection,”IEEE/ACM Trans. Audio, Speech, Lang. Process., vol. 28, pp. 2895–2905, 2020

    2020

  3. [3]

    From weak to strong sound event labels using adaptive change-point detection and active learning,

    J. Martinsson, O. Mogren, M. Sandsten, and T. Virtanen, “From weak to strong sound event labels using adaptive change-point detection and active learning,” inProc. 32nd Eur. Signal Process. Conf. (EUSIPCO), 2024, pp. 902–906

    2024

  4. [4]

    Active few-shot learning for sound event detection,

    Y . Wang, M. Cartwright, and J. P. Bello, “Active few-shot learning for sound event detection,” inProc. Interspeech, 2022, pp. 1551–1555

    2022

  5. [5]

    Active learning for sound event classification using monte-carlo dropout and PANN embeddings,

    S. Shishkin, D. Hollosi, S. Doclo, and S. Goetze, “Active learning for sound event classification using monte-carlo dropout and PANN embeddings,” inProc. Detection and Classification of Acoustic Scenes and Events Workshop (DCASE), 2021, pp. 150–154

    2021

  6. [6]

    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,” inProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), 2024, pp. 896–900

    2024

  7. [7]

    Online active learning for sound event detection,

    M. Lindsey, A. Shah, F. Kubala, and R. M. Stern, “Online active learning for sound event detection,”arXiv preprint arXiv:2309.14460, 2023

    arXiv 2023

  8. [8]

    Active few-shot learning for rare bioacoustic feature annotation,

    B. McEwen, K. Soltero, S. Gutschmidt, A. Bainbridge-Smith, J. Atlas, and R. Green, “Active few-shot learning for rare bioacoustic feature annotation,”Ecol. Inform., vol. 82, p. 102734, 2024

    2024

Show all 39 references
  1. [9]

    Aggregation strategies for efficient annotation of bioacoustic sound events using active learning,

    R. Lindholm, O. Marklund, O. Mogren, and J. Martinsson, “Aggregation strategies for efficient annotation of bioacoustic sound events using active learning,” inProc. 33rd Eur. Signal Process. Conf. (EUSIPCO), 2025, pp. 955–959

    2025

  2. [10]

    Reducing class imbalance during active learning for named entity annotation,

    K. Tomanek and U. Hahn, “Reducing class imbalance during active learning for named entity annotation,” inProc. 5th Int. Conf. Knowl. Capture, 2009, pp. 105–112

    2009

  3. [11]

    VaB-AL: Incorporating class imbalance and difficulty with variational Bayes for active learning,

    J. Choi, K. M. Yi, J. Kim, J. Choo, B. Kim, J. Chang, Y . Gwon, and H. J. Chang, “VaB-AL: Incorporating class imbalance and difficulty with variational Bayes for active learning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2021, pp. 6749–6758

    2021

  4. [12]

    Class-balanced loss based on effective number of samples,

    Y . Cui, M. Jia, T. Lin, Y . Song, and S. Belongie, “Class-balanced loss based on effective number of samples,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2019, pp. 9268–9277

    2019

  5. [13]

    Class- balanced active learning for image classification,

    J. Z. Bengar, J. van de Weijer, L. L. Fuentes, and B. Raducanu, “Class- balanced active learning for image classification,” inProc. IEEE/CVF Winter Conf. Appl. Comput. Vis. (WACV), 2022, pp. 3707–3716

    2022

  6. [14]

    Online adaptive asymmetric active learning for budgeted imbalanced data,

    Y . Zhang, P. Zhao, J. Cao, W. Ma, J. Huang, Q. Wu, and M. Tan, “Online adaptive asymmetric active learning for budgeted imbalanced data,” in Proc. 24th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining (KDD), 2018, pp. 2768–2777

    2018

  7. [15]

    Generative active learning for long-tailed instance segmentation,

    M. Zhu, C. Fan, H. Chen, Y . Liu, W. Mao, X. Xu, and C. Shen, “Generative active learning for long-tailed instance segmentation,” in Proc. 41st Int. Conf. Mach. Learn. (ICML), ser. Proc. Mach. Learn. Res., vol. 235, 2024, pp. 62 349–62 368

    2024

  8. [16]

    Deep long-tailed learning: A survey,

    Y . Zhang, B. Kang, B. Hooi, S. Yan, and J. Feng, “Deep long-tailed learning: A survey,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 45, no. 9, pp. 10 795–10 816, 2023

    2023

  9. [17]

    A sequential algorithm for training text classifiers,

    D. D. Lewis and W. A. Gale, “A sequential algorithm for training text classifiers,” inProc. 17th Annu. Int. ACM SIGIR Conf. Res. Develop. Inf. Retr.Springer, 1994, pp. 3–12

    1994

  10. [18]

    Settles,Active Learning, ser

    B. Settles,Active Learning, ser. Synth. Lect. Artif. Intell. Mach. Learn. Morgan and Claypool Publishers, 2012

    2012

  11. [19]

    Query by committee,

    H. S. Seung, M. Opper, and H. Sompolinsky, “Query by committee,” in Proc. 5th Annu. Workshop Comput. Learn. Theory, 1992, pp. 287–294

    1992

  12. [20]

    Selective sampling using the query by committee algorithm,

    Y . Freund, H. S. Seung, E. Shamir, and N. Tishby, “Selective sampling using the query by committee algorithm,”Mach. Learn., vol. 28, no. 2–3, pp. 133–168, 1997

    1997

  13. [21]

    Clustering to minimize the maximum intercluster distance,

    T. F. Gonzalez, “Clustering to minimize the maximum intercluster distance,”Theor. Comput. Sci., vol. 38, pp. 293–306, 1985

    1985

  14. [22]

    Active learning for convolutional neural networks: A core-set approach,

    O. Sener and S. Savarese, “Active learning for convolutional neural networks: A core-set approach,” inProc. Int. Conf. Learn. Represent. (ICLR), 2018

    2018

  15. [23]

    Deep Bayesian active learning with image data,

    Y . Gal, R. Islam, and Z. Ghahramani, “Deep Bayesian active learning with image data,” inProc. 34th Int. Conf. Mach. Learn. (ICML), ser. Proc. Mach. Learn. Res., vol. 70, 2017, pp. 1183–1192

    2017

  16. [24]

    The power of ensembles for active learning in image classification,

    W. H. Beluch, T. Genewein, A. N ¨urnberger, and J. M. K ¨ohler, “The power of ensembles for active learning in image classification,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), 2018, pp. 9368–9377

    2018

  17. [25]

    BatchBALD: Efficient and diverse batch acquisition for deep Bayesian active learning,

    A. Kirsch, J. van Amersfoort, and Y . Gal, “BatchBALD: Efficient and diverse batch acquisition for deep Bayesian active learning,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 32, 2019

    2019

  18. [26]

    Batch active learning at scale,

    G. Citovsky, G. DeSalvo, C. Gentile, L. Karydas, A. Rajagopalan, A. Rostamizadeh, and S. Kumar, “Batch active learning at scale,” inProc. Adv. Neural Inf. Process. Syst. (NeurIPS), vol. 34, 2021, pp. 11 933– 11 944

    2021

  19. [27]

    Batch active learning using determinantal point processes,

    E. Bıyık, K. Wang, N. Anari, and D. Sadigh, “Batch active learning using determinantal point processes,”arXiv preprint arXiv:1906.07975, 2019

    1906 arXiv

  20. [28]

    Deep batch active learning by diverse, uncertain gradient lower bounds,

    J. T. Ash, C. Zhang, A. Krishnamurthy, J. Langford, and A. Agarwal, “Deep batch active learning by diverse, uncertain gradient lower bounds,” inProc. Int. Conf. Learn. Represent. (ICLR), 2020

    2020

  21. [29]

    k-means++: the advantages of careful seeding,

    D. Arthur and S. Vassilvitskii, “k-means++: the advantages of careful seeding,” inProc. 18th Annu. ACM-SIAM Symp. Discrete Algorithms (SODA), 2007, pp. 1027–1035

    2007

  22. [30]

    Monte Carlo Markov chain algorithms for sampling strongly Rayleigh distributions and determinantal point processes,

    N. Anari, S. O. Gharan, and A. Rezaei, “Monte Carlo Markov chain algorithms for sampling strongly Rayleigh distributions and determinantal point processes,” inProc. 29th Annu. Conf. Learn. Theory, ser. Proc. Mach. Learn. Res., vol. 49, 2016, pp. 103–115

    2016

  23. [31]

    Determinantal point processes for machine learning,

    A. Kulesza and B. Taskar, “Determinantal point processes for machine learning,”Found. Trends Mach. Learn., vol. 5, no. 2–3, pp. 123–286, 2012

    2012

  24. [32]

    Near-optimal sensor placements in Gaussian processes: Theory, efficient algorithms and empirical studies,

    A. Krause, A. Singh, and C. Guestrin, “Near-optimal sensor placements in Gaussian processes: Theory, efficient algorithms and empirical studies,” J. Mach. Learn. Res., vol. 9, no. 8, pp. 235–284, 2008

    2008

  25. [33]

    An analysis of approximations for maximizing submodular set functions - I,

    G. L. Nemhauser, L. A. Wolsey, and M. L. Fisher, “An analysis of approximations for maximizing submodular set functions - I,”Math. Program., vol. 14, no. 1, pp. 265–294, 1978

    1978

  26. [34]

    HyenaSET: Hyena sound event transcripts and benchmark animal2vec performance for parsing animal communication,

    J. M. Woerner, C. Angonin, A. S. Gersick, K. E. Holekamp, F. H. Jensen, M. P. Johnson, M. H. M. Onsare, M. O. Pioon, J. C. Sch¨afer-Zimmermann, A. Strandburg-Peshkin, and E. D. Strauss, “HyenaSET: Hyena sound event transcripts and benchmark animal2vec performance for parsing a...

    2026

  27. [35]

    animal2vec and meerkat: A self-supervised transformer for rare-event raw audio input and a large-scale reference dataset for bioacoustics,

    J. C. Sch ¨afer-Zimmermann, V . Demartsev, B. Averly, K. L. Dhanjal- Adams, M. Duteil, G. Gall, M. Faiß, L. Johnson-Ulrich, D. Stowell, M. B. Manser, M. A. Roch, and A. Strandburg-Peshkin, “animal2vec and meerkat: A self-supervised transformer for rare-event raw audio input an...

    2026

  28. [36]

    An active learning method using clustering and committee-based sample selection for sound event classification,

    Z. Shuyang, T. Heittola, and T. Virtanen, “An active learning method using clustering and committee-based sample selection for sound event classification,” inProc. 16th Int. Workshop Acoust. Signal Enhancement (IWAENC), 2018, pp. 116–120

    2018

  29. [37]

    Hybrid disagreement-diversity active learning for bioacoustic sound event detection,

    S. Zhang and T. Virtanen, “Hybrid disagreement-diversity active learning for bioacoustic sound event detection,” inProc. 33rd Eur. Signal Process. Conf. (EUSIPCO), 2025, pp. 131–135

    2025

  30. [38]

    Submodularity in data subset selection and active learning,

    K. Wei, R. Iyer, and J. Bilmes, “Submodularity in data subset selection and active learning,” inProc. 32nd Int. Conf. Mach. Learn. (ICML), ser. Proc. Mach. Learn. Res., vol. 37, 2015, pp. 1954–1963

    2015

  31. [39]

    A simple sequentially rejective multiple test procedure,

    S. Holm, “A simple sequentially rejective multiple test procedure,”Scand. J. Stat., vol. 6, no. 2, pp. 65–70, 1979

    1979

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.