Pith. sign in

REVIEW 3 major objections 5 minor 50 references

For joint bird species and call-type classification, the best loss balancer depends on the backbone, how deep you adapt it, and which task you care about; freezing the backbone often beats full fine-tuning.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-12 03:23 UTC pith:YGH4BHRB

load-bearing objection Solid empirical map of loss balancing × backbone × probing on dual-label bird audio; the MTL-vs-STL ranking is overstated because the label spaces are not comparable. the 3 major comments →

arxiv 2607.03304 v1 pith:YGH4BHRB submitted 2026-07-03 cs.SD cs.LGeess.ASq-bio.QM

Adaptive Loss Balancing for Multi-Task Bioacoustic Classification of Bird Species and Call Types

Paria Vali Zadeh , Sven Tomforde This is my paper
classification cs.SD cs.LGeess.ASq-bio.QM
keywords BioacousticsPassive Acoustic MonitoringMulti-Task LearningAdaptive Loss BalancingBird Species ClassificationCall-Type ClassificationTransfer LearningLong-tailed Data
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Passive acoustic monitoring of birds needs models that can name both the species and the kind of call, yet those two labels are long-tailed and unevenly hard. This paper extends a joint multi-task setup on the WiWa forest-soundscape benchmark and systematically compares a fixed weighted loss against three adaptive balancers—uncertainty weighting, Dynamic Weight Averaging, and GradNorm—across four pretrained bird-domain encoders and three adaptation depths: linear probing, attentive probing, and full fine-tuning. The central finding is that there is no universal winner: factorising the two tasks helps call-type recognition most consistently, adaptive weighting helps species recognition more than call-type, and the preferred balancer and adaptation regime change with backbone and target. Full fine-tuning is not always best; frozen-backbone probing can be both more accurate and cheaper. The practical message is that loss balancing must be chosen together with the representation and the adaptation depth rather than treated as a free-standing hyperparameter.

Core claim

The preferred loss-balancing strategy for joint bird species and call-type classification depends on the backbone, the adaptation regime, and the target task, while frozen-backbone adaptation can deliver a better performance–efficiency trade-off than end-to-end fine-tuning. The factorised multi-task formulation yields the most consistent gains over a combined single-task baseline for call-type recognition; its effect on species recognition is regime-dependent.

What carries the argument

A shared pretrained bird-domain encoder with two separate classification heads whose binary cross-entropy losses are combined by either a fixed weight, homoscedastic uncertainty weighting, Dynamic Weight Averaging, or GradNorm, evaluated under linear probing, attentive probing, and full fine-tuning.

Load-bearing premise

That class-masked average precision on the WiWa test segments remains a fair comparison even though the multi-task and combined single-task models operate on different label spaces.

What would settle it

On the same WiWa test-5s partition, recompute class-masked cmAP after forcing multi-task and combined single-task models onto identical output spaces (or after excluding all classes that never appear in the test set); if the ranking of multi-task over single-task and of frozen over full fine-tuning reverses for the leading backbones, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

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 / 5 minor

Summary. The paper extends BirdCallNet to joint species and call-type classification on the long-tailed WiWa forest-soundscape benchmark. It evaluates four bird-domain pretrained encoders (ConvNeXtBS, EAT, BirdMAE, ProtoCLR) under linear probing, attentive probing, and full fine-tuning, comparing a fixed weighted-sum multi-task objective against homoscedastic uncertainty weighting, Dynamic Weight Averaging, and (under FT only) GradNorm. The central claims are that factorised multi-task learning yields the most consistent gains over a combined single-task baseline for call-type recognition, that full fine-tuning is not uniformly optimal, that adaptive weighting helps species recognition more consistently than call-type recognition, and that the preferred balancing strategy depends on backbone, adaptation regime, and target task, with frozen-backbone adaptation often offering a favourable performance–efficiency trade-off.

Significance. If the comparative findings hold under matched evaluation, the work is a useful, carefully controlled empirical contribution to multi-task bioacoustics. It systematically maps how standard adaptive loss-balancing methods interact with modern bird-domain representations and probing regimes on a dual-label, long-tailed soundscape benchmark—an under-studied setting relative to pure species ID. Strengths include multi-seed reporting, explicit regime–strategy applicability (GradNorm restricted to FT), transparent hyperparameter tables, and an honest limitations discussion. The practical takeaway that frozen adaptation can outperform end-to-end fine-tuning for some backbones is actionable for passive acoustic monitoring pipelines. The study is incremental relative to the conference BirdCallNet version but substantially expands the optimisation analysis.

major comments (3)
  1. Table 3 and §6.1 (also Abstract and §7): The claim that factorised multi-task learning yields the most consistent improvements over the combined single-task baseline for call-type recognition is not metric-equivalent. STL optimises and evaluates cmAP over the joint species#call-type vocabulary, while MTL reports separate class-masked cmAP on the call-type head alone (§5.1, §5.6). The paper itself notes the formulations “operate on different label spaces” and supply only “complementary evidence” (§6.6), yet still ranks MTL as superior. Because the combined space is far sparser and longer-tailed (561 observed combinations, median 26; Fig. 2c, Fig. 3), STL cmAP is structurally harder; apparent MTL gains may be an artifact of the easier marginal call-type space. A matched evaluation (e.g., projecting STL predictions onto the call-type marginal, or reporting joint metrics for both formulation
  2. Table 1 and Table 4: GradNorm under FT uses batch size 8 (vs 128 for all other strategies) because of per-task gradient memory overhead. This confounds the FT comparison: any under-performance of GradNorm on species (and its mixed call-type results) cannot be cleanly attributed to the weighting mechanism versus the reduced batch size and altered optimisation dynamics. Either re-run GradNorm at matched batch size (or with gradient accumulation) or explicitly qualify that the GradNorm numbers are not directly comparable and remove them from the main ranking statements.
  3. §5.6 and §6.6: Class-masked cmAP retains the original output space and assigns zero AP to classes without positive test examples. The paper acknowledges this can depress the macro-average, especially for call types. Given that the strongest claims rest on small absolute differences across weighting strategies (often <0.02 cmAP), the evaluation protocol should either exclude undefined classes from the average or report both masked and unmasked variants so that ranking stability can be assessed.
minor comments (5)
  1. Table 3: Score is defined as the mean of Naive/Unc./DWA only; GradNorm is excluded. This is reasonable given applicability, but the caption should state it explicitly so readers do not misread the FT Score column.
  2. §3.2 / Eq. (3)–(11): The fixed weights λ_sp=1.0, λ_ct=17.0 are carried over from the conference version without re-tuning under the new adaptive baselines. A short sensitivity check (or statement that they remain near-optimal) would strengthen the Naive baseline.
  3. Fig. 1 caption and §5.2: Attentive probing parameter count is given as 2d²+(C+1)d+C; a brief note on whether multi-head attention uses a single shared query or class-specific queries would aid reproducibility.
  4. Throughout: minor typographic inconsistencies (e.g., “multi-task” vs “multi-task”, “ConvNextBS” vs “ConvNeXtBS” in Table 3 header) should be normalised.
  5. §4.2: The train–test asymmetry (focal-style training clips vs multi-label soundscape test) is well described; adding a one-sentence reminder in the results discussion would help readers interpret absolute cmAP levels.

Circularity Check

0 steps flagged

Empirical multi-task benchmarking paper; results are measured on held-out data with standard adaptive-weighting formulas, not derived by construction from inputs.

full rationale

The paper is an experimental comparison of fixed and adaptive loss-balancing strategies (uncertainty weighting, DWA, GradNorm) across four pretrained encoders and three adaptation regimes on the WiWa benchmark. Claims about relative performance (e.g., factorised MTL vs combined STL for call-type, preferred weighting by regime) rest on class-masked cmAP tables computed from independent training runs on held-out test-5s segments. Fixed weights are taken from prior validation tuning and applied as a baseline; adaptive methods use published, parameter-light formulas that do not embed the target metrics. Mild self-extension from the authors' conference BirdCallNet paper supplies the dataset characterisation, architecture, and two of three seeds for the Naive/STL baselines, but the adaptive-weighting experiments and all new comparisons are freshly measured. No equation equates a claimed prediction to a fitted input, no uniqueness theorem is imported to force the result, and no ansatz is smuggled via self-citation. Label-space mismatch between STL and MTL is a fairness/correctness concern (acknowledged in §6.6) rather than circularity. Score 1 reflects only the ordinary self-extension of a conference-to-journal paper; the central empirical rankings do not reduce by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

This is an empirical multi-task transfer study. The central claims rest on standard MTL loss-balancing formulas, pretrained BirdSet-domain encoders, the WiWa dual-label construction, and several hand-chosen optimisation settings rather than new physical entities or first-principles axioms.

free parameters (5)
  • Fixed task weights (lambda_sp, lambda_ct) = 1.0 / 17.0
    Manually validation-tuned coefficients 1.0 and 17.0 from the conference version used as the Naive baseline; central comparisons depend on this choice.
  • DWA temperature T = 2.0
    Set to 2.0 following the original DWA paper and empirical preference; controls weight sharpness.
  • GradNorm asymmetry alpha = 1.5
    Set to 1.5 following the original GradNorm configuration; controls training-rate balancing strength.
  • Learning rates and early-stopping deltas per regime = see Table 1
    Regime-specific LRs (5e-3 / 1.3e-3 / 1e-4) and STL/MTL min-delta thresholds are chosen settings that affect which checkpoint is selected.
  • GradNorm FT batch size = 8
    Reduced to 8 for memory reasons, unlike batch 128 elsewhere; can change optimisation dynamics in the GradNorm comparison.
axioms (5)
  • domain assumption Hard-parameter-sharing multi-task learning with separate species and call-type heads is an appropriate model of joint bioacoustic annotation.
    Section 3.2 formulates the architecture and losses under this standard MTL assumption.
  • domain assumption Binary cross-entropy with logits on multi-hot labels is a suitable objective for segment-level multi-label evaluation.
    Used for both heads and the combined baseline (Sections 3.2, 5.1).
  • domain assumption Class-masked macro average precision (cmAP) is the primary figure of merit under long-tailed labels.
    Section 5.6 defines evaluation exclusively via cmAP on classes present in the eval split.
  • standard math Homoscedastic uncertainty weighting, DWA, and GradNorm are valid adaptive balancers for two-task audio classification.
    Equations (4)-(12) import standard published formulations without re-derivation.
  • domain assumption WiWa train/test construction after label harmonisation is a valid dual-label passive-monitoring benchmark despite train-test domain asymmetry.
    Section 4 motivates and constructs the benchmark from focal-like training clips and soundscape test segments.
invented entities (1)
  • BirdCallNet multi-task factorisation on WiWa no independent evidence
    purpose: Separate species and call-type heads over shared bird-domain encoders for joint classification.
    Named extension of the authors' prior conference system; not a new physical entity, but a paper-specific experimental construct.

pith-pipeline@v1.1.0-grok45 · 26913 in / 3427 out tokens · 24174 ms · 2026-07-12T03:23:29.085243+00:00 · methodology

0 comments
read the original abstract

Reliable analysis of bird vocalisations in passive acoustic monitoring requires models handling multiple, imbalanced annotation targets. We extend BirdCallNet for joint species and call-type classification on the long-tailed WiWa dataset and investigate how task-loss balancing interacts with pretrained representations and adaptation depth. We evaluate four bird-domain encoders, ConvNeXtBS, EAT, BirdMAE, and ProtoCLR, with separate species and call-type heads under linear probing, attentive probing, and full fine-tuning. A manually tuned fixed objective is compared with homoscedastic uncertainty weighting and Dynamic Weight Averaging across all three adaptation regimes, while GradNorm is evaluated only under full fine-tuning. Results indicate that the factorised multi-task formulation yields the most consistent improvements over the combined single-task baseline for call-type recognition, while its effect on species recognition depends on the adaptation regime. Full fine-tuning is not consistently optimal: ConvNeXtBS achieves the highest mean species performance under linear probing, whereas BirdMAE provides the strongest call-type performance under attentive probing. Adaptive weighting benefits species recognition more consistently than call-type recognition. Uncertainty weighting is particularly effective for species recognition under attentive probing, whereas Dynamic Weight Averaging is generally stronger for the same task under full fine-tuning. GradNorm achieves competitive call-type performance for selected backbones but consistently underperforms other weighting strategies for species recognition and incurs higher computational and memory costs. Overall, the preferred loss-balancing strategy depends on the backbone, adaptation regime, and target task, while frozen-backbone adaptation can provide a more favourable performance-efficiency trade-off than end-to-end fine-tuning.

Figures

Figures reproduced from arXiv: 2607.03304 by Paria Vali Zadeh, Sven Tomforde.

Figure 1
Figure 1. Figure 1: Schematic overview of the proposed multi-task learning (MTL) framework for bird vocalisation analysis. Input audio segments are preprocessed and passed through one of the pretrained audio backbones. The shared representation is kept frozen for linear probing (LP) and attentive probing (AP), whereas it is updated during full fine-tuning (FT). Species and call-type predictions are optimised through task-spec… view at source ↗
Figure 2
Figure 2. Figure 2: Label imbalance in the processed WiWa training split: (a) species-level long￾tail distribution; (b) biological call-type distribution on a logarithmic scale; and (c) combined-label long-tail distribution over biological species–call-type pairs. The nonbird background class is excluded from these biological analyses [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Species–call-type co-occurrence matrix for the processed WiWa training split. Rows show the 60 biological species with the largest number of samples, and columns show the ten biological call types. Cell colours indicate ln(1 + count). The nonbird background class is excluded. Across the full biological matrix, 561 of 2,040 possible species–call-type cells are observed, leaving 72.5% unobserved. architectur… view at source ↗
Figure 4
Figure 4. Figure 4: Distribution of sufficiently represented biological call types per species in the processed WiWa training split. For each species, we count the number of call types with more than 25, 50, and 100 samples. The nonbird background class is excluded. 5.1 Task Formulation and Baseline We treat the WiWa benchmark (see Section 4.2) as a segment-level multi-label classification problem. The model employs two indep… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

50 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    https://doi.org/10.48550/arXiv.1610.01644, http://arxiv.org/ abs/1610.01644, arXiv:1610.01644 [stat]

    Alain, G., Bengio, Y.: Understanding intermediate layers using linear classifier probes (Nov 2018). https://doi.org/10.48550/arXiv.1610.01644, http://arxiv.org/ abs/1610.01644, arXiv:1610.01644 [stat]

    Pith/arXiv arXiv doi:10.48550/arxiv.1610.01644 2018

  2. [2]

    Ecological Informatics27, 21–25 (May 2015)

    Arriaga, J.G., Cody, M.L., Vallejo, E.E., Taylor, C.E.: Bird-DB: A database for anno- tated bird song sequences. Ecological Informatics27, 21–25 (May 2015). https://doi. org/10.1016/j.ecoinf.2015.01.007, https://www.sciencedirect.com/science/article/ pii/S1574954115000151

    work page doi:10.1016/j.ecoinf.2015.01.007 2015

  3. [3]

    In: Pro- ceedings of the 39th International Conference on Machine Learning

    Baevski, A., Hsu, W.N., Xu, Q., Babu, A., Gu, J., Auli, M.: data2vec: A General Framework for Self-supervised Learning in Speech, Vision and Language. In: Pro- ceedings of the 39th International Conference on Machine Learning. pp. 1298–1312. PMLR (Jun 2022), https://proceedings.mlr.press/v162/baevski22a.html

    2022

  4. [4]

    Scientific Reports11(1), 15733 (Aug 2021)

    Bravo Sanchez, F.J., Hossain, M.R., English, N.B., Moore, S.T.: Bioacoustic classi- fication of avian calls from raw sound waveforms with an open-source deep learning architecture. Scientific Reports11(1), 15733 (Aug 2021). https://doi.org/10.1038/ s41598-021-95076-6, https://www.nature.com/articles/s41598-021-95076-6

    2021

  5. [5]

    Conservation Biology39(5), e70058 (2025)

    Brunk, K.M., Kramer, H.A., Peery, M.Z., Kahl, S., Wood, C.M.: Assess- ing spatial variability and efficacy of surrogate species at an ecosystem scale. Conservation Biology39(5), e70058 (2025). https://doi.org/10.1111/ cobi.70058, https://onlinelibrary.wiley.com/doi/abs/10.1111/cobi.70058, _eprint: https://conbio.onlinelibrary.wiley.com/doi/pdf/10.1111/cobi.70058

    work page doi:10.1111/cobi.70058 2025

  6. [6]

    Machine Learning28(1), 41–75 (Jul 1997)

    Caruana, R.: Multitask Learning. Machine Learning28(1), 41–75 (Jul 1997). https: //doi.org/10.1023/A:1007379606734, https://doi.org/10.1023/A:1007379606734

    doi:10.1023/a:1007379606734 1997

  7. [7]

    https://doi.org/10.48550/ arXiv.2212.09058, http://arxiv.org/abs/2212.09058, arXiv:2212.09058 [eess]

    Chen, S., Wu, Y., Wang, C., Liu, S., Tompkins, D., Chen, Z., Wei, F.: BEATs: Audio Pre-Training with Acoustic Tokenizers (Dec 2022). https://doi.org/10.48550/ arXiv.2212.09058, http://arxiv.org/abs/2212.09058, arXiv:2212.09058 [eess]

    Pith/arXiv arXiv 2022

  8. [8]

    https://doi.org/10.48550/arXiv.1711.02257, http://arxiv.org/abs/1711

    Chen, Z., Badrinarayanan, V., Lee, C.Y., Rabinovich, A.: GradNorm: Gradi- ent Normalization for Adaptive Loss Balancing in Deep Multitask Networks (Jun 2018). https://doi.org/10.48550/arXiv.1711.02257, http://arxiv.org/abs/1711. 02257, arXiv:1711.02257 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.02257 2018

  9. [9]

    Proceedings of the Royal Society B: Biological Sciences288(1947), 20210225 (Mar 2021)

    Crates, R., Langmore, N., Ranjard, L., Stojanovic, D., Rayner, L., Ingwersen, D., Heinsohn, R.: Loss of vocal culture and fitness costs in a critically endangered song- bird. Proceedings of the Royal Society B: Biological Sciences288(1947), 20210225 (Mar 2021). https://doi.org/10.1098/rspb.2021.0225, https://doi.org/10.1098/rspb. 2021.0225

    work page doi:10.1098/rspb.2021.0225 1947

  10. [10]

    https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (Jun 2021). https://doi.org/10.48550/arXiv.2010.11929, http://arxiv.org/abs/2010. 11929, arXiv:2010.11929 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2010.11929 2021

  11. [11]

    Annals of the New York Academy Adaptive Loss Balancing for Bioacoustic Classification 27 of Sciences1134(1), 233–266 (2008)

    Drewitt, A.L., Langston, R.H.: Collision Effects of Wind-power Genera- tors and Other Obstacles on Birds. Annals of the New York Academy Adaptive Loss Balancing for Bioacoustic Classification 27 of Sciences1134(1), 233–266 (2008). https://doi.org/10.1196/annals.1439. 015, https://onlinelibrary.wiley.com/doi/abs/10.1196/annals.1439.015, _eprint: https://ny...

    work page doi:10.1196/annals.1439 2008

  12. [12]

    https://doi.org/10.48550/arXiv.2401.08541, http://arxiv.org/abs/2401

    El-Nouby, A., Klein, M., Zhai, S., Bautista, M.A., Toshev, A., Shankar, V., Susskind, J.M., Joulin, A.: Scalable Pre-training of Large Autoregressive Image Models (Jan 2024). https://doi.org/10.48550/arXiv.2401.08541, http://arxiv.org/abs/2401. 08541, arXiv:2401.08541 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2401.08541 2024

  13. [13]

    https://doi.org/10.48550/arXiv.2204.11479, http://arxiv.org/ abs/2204.11479, arXiv:2204.11479 [cs]

    Gazneli, A., Zimerman, G., Ridnik, T., Sharir, G., Noy, A.: End-to-End Audio Strikes Back: Boosting Augmentations Towards An Efficient Audio Classification Network (Jul 2022). https://doi.org/10.48550/arXiv.2204.11479, http://arxiv.org/ abs/2204.11479, arXiv:2204.11479 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2204.11479 2022

  14. [14]

    Scientific Reports13(1), 22876 (Dec 2023)

    Ghani, B., Denton, T., Kahl, S., Klinck, H.: Global birdsong embeddings enable su- perior transfer learning for bioacoustic classification. Scientific Reports13(1), 22876 (Dec 2023). https://doi.org/10.1038/s41598-023-49989-z, https://www.nature.com/ articles/s41598-023-49989-z

    doi:10.1038/s41598-023-49989-z 2023

  15. [15]

    Methods in Ecology and Evolution10(2), 169–185 (2019)

    Gibb, R., Browning, E., Glover-Kapfer, P., Jones, K.E.: Emerging opportunities and challenges for passive acoustics in ecological assessment and monitoring. Methods in Ecology and Evolution10(2), 169–185 (2019). https://doi.org/10.1111/2041-210X. 13101, https://onlinelibrary.wiley.com/doi/abs/10.1111/2041-210X.13101, _eprint: https://besjournals.onlinelib...

    work page doi:10.1111/2041-210x 2019

  16. [16]

    https://doi.org/10.48550/arXiv.2104.01778, http://arxiv.org/abs/2104

    Gong, Y., Chung, Y.A., Glass, J.: AST: Audio Spectrogram Transformer (Jul 2021). https://doi.org/10.48550/arXiv.2104.01778, http://arxiv.org/abs/2104. 01778, arXiv:2104.01778 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.2104.01778 2021

  17. [17]

    In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)

    Hagiwara, M.: AVES: Animal Vocalization Encoder Based on Self-Supervision. In: ICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). pp. 1–5 (Jun 2023). https://doi.org/10.1109/ICASSP49357. 2023.10095642, https://ieeexplore.ieee.org/abstract/document/10095642

    doi:10.1109/icassp49357 2023

  18. [18]

    Advances in Neural Information Processing Systems35, 28708–28720 (Dec 2022), https://proceedings.neurips

    Huang, P.Y., Xu, H., Li, J., Baevski, A., Auli, M., Galuba, W., Metze, F., Feicht- enhofer, C.: Masked Autoencoders that Listen. Advances in Neural Information Processing Systems35, 28708–28720 (Dec 2022), https://proceedings.neurips. cc/paper_files/paper/2022/hash/b89d5e209990b19e33b418e14f323998-Abstract- Conference.html

    2022

  19. [19]

    In: Ralph, C

    Ii, A.M.M.: Bird strike and electrocutions at power lines, communication towers, and wind turbines: state of the art and state of the science - next steps toward mitigation. In: Ralph, C. John; Rich, Terrell D., editors 2005. Bird Conservation Implementation and Integration in the Americas: Proceedings of the Third International Partners in Flight Confere...

    2005

  20. [20]

    Kahl, S., Denton, T., Klinck, H., Reers, H., Cherutich, F., Glotin, H., Goeau, H., Vellinga, W.P., Planqué, R., Joly, A.: Overview of BirdCLEF 2023: Automated bird species identification in Eastern Africa (2023), https://agritrop.cirad.fr/611094/

    2023

  21. [21]

    Ecological Informatics61, 101236 (Mar 2021)

    Kahl, S., Wood, C.M., Eibl, M., Klinck, H.: BirdNET: A deep learning solution for avian diversity monitoring. Ecological Informatics61, 101236 (Mar 2021). https: //doi.org/10.1016/j.ecoinf.2021.101236, https://www.sciencedirect.com/science/ article/pii/S1574954121000273

    doi:10.1016/j.ecoinf.2021.101236 2021

  22. [22]

    In: Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR)

    Kendall, A., Gal, Y., Cipolla, R.: Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In: Proceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR). pp. 7482–7491 (Jun 28 P. Vali Zadeh and S. Tomforde 2018), https://openaccess.thecvf.com/content_cvpr_2018/html/Kendall_Multi- Task_Learning_U...

    2018

  23. [23]

    In: Proceedings of the 2019 7th International Conference on Information Technology: IoT and Smart City

    Kim, D., Lee, Y., Ko, H.: Multi-task Learning for Animal Species and Group Category Classification. In: Proceedings of the 2019 7th International Conference on Information Technology: IoT and Smart City. pp. 435–438. ICIT ’19, Association for Computing Machinery, New York, NY, USA (Mar 2020). https://doi.org/10. 1145/3377170.3377259, https://dl.acm.org/do...

    work page doi:10.1145/3377170.3377259 2019

  24. [24]

    Conserva- tion Biology35(1), 50–63 (2021)

    Lewis, R.N., Williams, L.J., Gilman, R.T.: The uses and impli- cations of avian vocalizations for conservation planning. Conserva- tion Biology35(1), 50–63 (2021). https://doi.org/10.1111/cobi.13465, https://onlinelibrary.wiley.com/doi/abs/10.1111/cobi.13465, _eprint: https://conbio.onlinelibrary.wiley.com/doi/pdf/10.1111/cobi.13465

    work page doi:10.1111/cobi.13465 2021

  25. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, S., Johns, E., Davison, A.J.: End-to-end multi-task learning with attention. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 1871–1880 (Jun 2019), https://openaccess.thecvf.com/content_CVPR_2019/html/Liu_End-To- End_Multi-Task_Learning_With_Attention_CVPR_2019_paper.html

    2019

  26. [26]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Liu, Z., Mao, H., Wu, C.Y., Feichtenhofer, C., Darrell, T., Xie, S.: A ConvNet for the 2020s. In: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11966–11976 (Jun 2022). https://doi.org/10.1109/CVPR52688.2022. 01167, https://ieeexplore.ieee.org/document/9879745

    doi:10.1109/cvpr52688.2022 2022

  27. [27]

    https://doi.org/10.48550/arXiv.1711.05101, http://arxiv.org/abs/1711

    Loshchilov, I., Hutter, F.: Decoupled Weight Decay Regularization (Jan 2019). https://doi.org/10.48550/arXiv.1711.05101, http://arxiv.org/abs/1711. 05101, arXiv:1711.05101 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv.1711.05101 2019

  28. [28]

    The Condor: Ornithological Applications116(1), 8–23 (Feb 2014)

    Loss, S.R., Will, T., Loss, S.S., Marra, P.P.: Bird–building collisions in the United States: Estimates of annual mortality and species vulnerability. The Condor: Ornithological Applications116(1), 8–23 (Feb 2014). https://doi.org/10.1650/ CONDOR-13-090.1, https://doi.org/10.1650/CONDOR-13-090.1

    work page doi:10.1650/condor-13-090.1 2014

  29. [29]

    https://doi.org/10.5281/zenodo.1208080, https://zenodo.org/records/1208080

    Lostanlen, V., Salamon, J., Farnsworth, A., Kelling, S., Bello, J.P.: BirdVox- DCASE-20k: a dataset for bird audio detection in 10-second clips (Mar 2018). https://doi.org/10.5281/zenodo.1208080, https://zenodo.org/records/1208080

    work page doi:10.5281/zenodo.1208080 2018

  30. [30]

    https://www

    Macaulay Library, Cornell Lab of Ornithology: Macaulay library. https://www. macaulaylibrary.org/, last accessed 2 July 2026

    2026

  31. [31]

    Elsevier (Oct 2004), google-Books-ID: 2iFmsVSyV4gC

    Marler, P.R., Slabbekoorn, H.: Nature’s Music: The Science of Birdsong. Elsevier (Oct 2004), google-Books-ID: 2iFmsVSyV4gC

    2004

  32. [32]

    https://doi.org/10.5281/zenodo.20762099, https://zenodo.org/records/20762099

    Martin, R., Reers, H., Abraham, M., Brauns, K., Haupt, J., Heinrich, R., Kahl, S., Raabe, J., Rauch, L., Schäfermeier, B., Scholz, C., Schwinger, R., Sick, B., Tomforde, S., Vali Zadeh, P., Wirth, M.: CEB: A collection of strongly-labeled bird soundscapes and Xeno-Canto recordings from Central Europe, with bounding boxes and vocalization types (Jun 2026)....

    work page doi:10.5281/zenodo.20762099 2026

  33. [33]

    Frontiers in Bird Science3(Jul 2024)

    van Merriënboer, B., Hamer, J., Dumoulin, V., Triantafillou, E., Denton, T.: Birds, bats and beyond: evaluating generalization in bioacoustics models. Frontiers in Bird Science3(Jul 2024). https://doi.org/10.3389/fbirs.2024.1369756, https://www. frontiersin.org/journals/bird-science/articles/10.3389/fbirs.2024.1369756/full

    doi:10.3389/fbirs.2024.1369756 2024

  34. [34]

    Miron, M., Robinson, D., Alizadeh, M., Gilsenan-McMahon, E., Narula, G., Chemla, E., Cusimano, M., Effenberger, F., Hagiwara, M., Hoffman, B., Keen, S., Kim, D., Lawton, J., Liu, J.Y., Raskin, A., Pietquin, O., Geist, M.: What Matters for Bioacoustic Encoding (Aug 2025), https://arxiv.org/abs/2508.11845v2 Adaptive Loss Balancing for Bioacoustic Classification 29

    Pith/arXiv arXiv 2025

  35. [35]

    PeerJ Computer Science5, e223 (Oct 2019)

    Morfi, V., Bas, Y., Pamuła, H., Glotin, H., Stowell, D.: NIPS4Bplus: a richly annotated birdsong audio dataset. PeerJ Computer Science5, e223 (Oct 2019). https://doi.org/10.7717/peerj-cs.223, https://peerj.com/articles/cs-223

    work page doi:10.7717/peerj-cs.223 2019

  36. [36]

    Moummad, I., Serizel, R., Benetos, E., Farrugia, N.: Domain-Invariant Representa- tion Learning of Bird Sounds (Sep 2024), https://arxiv.org/abs/2409.08589v7

    arXiv 2024

  37. [37]

    https://doi.org/10.48550/arXiv

    Rauch, L., Heinrich, R., Moummad, I., Joly, A., Sick, B., Scholz, C.: Can Masked Autoencoders Also Listen to Birds? (Aug 2025). https://doi.org/10.48550/arXiv. 2504.12880, http://arxiv.org/abs/2504.12880, arXiv:2504.12880 [cs]

    Pith/arXiv arXiv doi:10.48550/arxiv 2025

  38. [38]

    Rauch, L., Schwinger, R., Wirth, M., Heinrich, R., Huseljic, D., Herde, M., Lange, J., Kahl, S., Sick, B., Tomforde, S., Scholz, C.: BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics. International Conference on Learning Repre- sentations2025, 29482–29520 (May 2025), https://proceedings.iclr.cc/paper_files/ paper/2025/hash/484d...

    2025

  39. [39]

    Reichenbach, M., Reers, H., Günther, F., Menke, K., Grimm, J., Martin, R.: Auswirkungen von WEA auf die akustische Aktivität ausgewählter Waldvogelarten: Untersuchungen zu Verdrängungseffekten mittels automatisierter Erfassung: Ergeb- nisse des F+E-Vorhabens: Optimierung des Planungs- und Genehmigungsprozesses von Windenergieanlagen im Wald hinsichtlich B...

    work page doi:10.19217/skr643 2022

  40. [40]

    Science366(6461), 120–124 (Oct 2019)

    Rosenberg, K.V., Dokter, A.M., Blancher, P.J., Sauer, J.R., Smith, A.C., Smith, P.A.,Stanton,J.C.,Panjabi,A.,Helft,L.,Parr,M.,Marra,P.P.:DeclineoftheNorth American avifauna. Science366(6461), 120–124 (Oct 2019). https://doi.org/10. 1126/science.aaw1313, https://www.science.org/doi/full/10.1126/science.aaw1313

    work page doi:10.1126/science.aaw1313 2019

  41. [41]

    Functional Ecology37(4), 959–975 (2023)

    Ross, S.R.P.J., O’Connell, D.P., Deichmann, J.L., Desjonquères, C., Gasc, A., Phillips, J.N., Sethi, S.S., Wood, C.M., Burivalova, Z.: Passive acoustic monitoring provides a fresh perspective on fundamental ecological questions. Functional Ecology37(4), 959–975 (2023). https://doi.org/10.1111/1365-2435. 14275, https://onlinelibrary.wiley.com/doi/abs/10.11...

    work page doi:10.1111/1365-2435 2023

  42. [42]

    https://doi.org/10.48550/arXiv.2508.01277, http://arxiv.org/abs/2508

    Schwinger, R., Zadeh, P.V., Rauch, L., Kurz, M., Hauschild, T., Lapp, S., Tomforde, S.: Foundation Models for Bioacoustics – a Comparative Review (Aug 2025). https://doi.org/10.48550/arXiv.2508.01277, http://arxiv.org/abs/2508. 01277, arXiv:2508.01277 [cs]

    doi:10.48550/arxiv.2508.01277 2025

  43. [43]

    Computational bioacoustics with deep learning: a review and roadmap

    Stowell, D.: Computational bioacoustics with deep learning: a review and roadmap. PeerJ10, e13152 (Mar 2022). https://doi.org/10.7717/peerj.13152, http://arxiv. org/abs/2112.06725, arXiv:2112.06725 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.7717/peerj.13152 2022

  44. [44]

    Conservation Science and Practice1(8), e72 (Aug 2019)

    Teixeira, D., Maron, M., Rensburg, B.J.v.: Bioacoustic monitoring of animal vocal behavior for conservation. Conservation Science and Practice1(8), e72 (Aug 2019). https://doi.org/10.1111/csp2.72, https://conbio.onlinelibrary.wiley.com/doi/ 10.1111/csp2.72

    doi:10.1111/csp2.72 2019

  45. [45]

    Nature Communications 13(1), 792 (Feb 2022)

    Tuia, D., Kellenberger, B., Beery, S., Costelloe, B.R., Zuffi, S., Risse, B., Mathis, A., Mathis, M.W., van Langevelde, F., Burghardt, T., Kays, R., Klinck, H., Wikelski, M., Couzin, I.D., van Horn, G., Crofoot, M.C., Stewart, C.V., Berger-Wolf, T.: Perspectives in machine learning for wildlife conservation. Nature Communications 13(1), 792 (Feb 2022). ht...

    doi:10.1038/s41467-022-27980-y 2022

  46. [46]

    Conservation Biol- ogy27(1), 210–218 (2013)

    Valderrama, S.V., Molles, L.E., Waas, J.R.: Effects of Population Size on Singing Behavior of a Rare Duetting Songbird. Conservation Biol- ogy27(1), 210–218 (2013). https://doi.org/10.1111/j.1523-1739.2012.01917.x, 30 P. Vali Zadeh and S. Tomforde https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1523-1739.2012.01917.x, _eprint: https://conbio.onlinelibra...

    work page doi:10.1111/j.1523-1739.2012.01917.x 2013

  47. [47]

    In: Proceedings of the 18th International Conference on Agents and Artificial Intelligence (ICAART)

    Vali Zadeh, P., Tomforde, S.: BirdCallNet: Joint species and call-type classification. In: Proceedings of the 18th International Conference on Agents and Artificial Intelligence (ICAART). pp. 4744–4753. SCITEPRESS, Marbella, Spain (2026). https://doi.org/10.5220/0014488900004052, https://www.scitepress.org/Link.aspx? doi=10.5220/0014488900004052

    work page doi:10.5220/0014488900004052 2026

  48. [48]

    In: CLEF (Working Notes) (2015)

    Vellinga, W.P., Planqué, R.: The xeno-canto collection and its relation to sound recognition and classification. In: CLEF (Working Notes) (2015)

    2015

  49. [49]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Wu, H., Xiao, B., Codella, N., Liu, M., Dai, X., Yuan, L., Zhang, L.: CvT: Introducing convolutions to vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 22–31 (Oct 2021), https://openaccess.thecvf.com/content/ICCV2021/html/Wu_CvT_ Introducing_Convolutions_to_Vision_Transformers_ICCV_2021_paper.html

    2021

  50. [50]

    IEEE Transactions on Knowledge and Data Engineering34(12), 5586–5609 (Dec 2022)

    Zhang, Y., Yang, Q.: A Survey on Multi-Task Learning. IEEE Transactions on Knowledge and Data Engineering34(12), 5586–5609 (Dec 2022). https://doi.org/ 10.1109/TKDE.2021.3070203, https://ieeexplore.ieee.org/document/9392366/

    doi:10.1109/tkde.2021.3070203 2022