Active Learning for Bioacoustics

Results

Task description

Task 4 challenges participants to develop an active learning sampling strategy for bioacoustic classification. Participants implement a sampling function within the BaseAL framework, which iteratively selects batches of samples, reveals their labels via oracle labelling, and retrains a classification head on the growing labelled set. Submissions are evaluated using the Area Under the Learning Curve (AULC) of mean average precision (mAP), aggregated across both a terrestrial dataset (BirdSet) and a marine dataset (ATBFL), requiring methods to generalise across domains.

Task description page

Awards

The official winning submission is determined by the official evaluation ranking. The Jury Award is selected by the task organisers based on the submitted systems and technical reports.

Official winning submission

Gabriel Dubus, Hugo Magaldi and Anatole Gros-Martial
Eco-Anthropologie, Mus\'eum National d'Histoire Naturelle, UMR7206, CNRS, Paris, France; Centre d'Etudes Biologiques de Chiz\'e, UMR 7372, CNRS, La Rochelle Universit\'e, France
Adaptive Diversity-Uncertainty Active Learning with Redundancy Control for Bioacoustic Event Classification

Jury award

Haoyu Wang, Han Dou, Hui Chen, Yushan Du, Liushi Tu, Ningning Pan, Hao Zhang, Jingyao Wu, Gang Li and Gongping Huang
Southwestern University of Finance and Economics, China; MiLM Plus, Xiaomi Inc.; IASP Lab, Wuhan University, China; Massachusetts Institute of Technology, USA
PB-MFS: Multi-Funnel Selection with Pareto-Balanced Uncertainty and Diversity for Bioacoustic Active Learning
Selected due to the novelty of their candidate filtering method and the quality and completeness of their technical report

Submission ranking

Submissions are ranked on the averaged AULC (mAP) across each of the data subsets. Supplementary metrics for computational and annotation efficiency are reported as well, but not used for ranking.

Submission information Rank Efficiency metrics
Corresponding Author System name Technical
report		 Official
rank		 Rank
score		 Computational
cost (rel.)		 Sampling
time (s)		 Annotation
cost
Gabriel Dubus ADU-MMR dubusAdaptive2026 1 0.507 2.000 1.006 1026.700
Hugo Magaldi CARE-DPP magaldiCARE2026 2 0.505 1.100 0.215 1124.550
Haoyu Wang PB-MFS wangPBMFS2026 3 0.499 1.600 0.485 1109.900
Wayne Yang Safe Rarity yangSafeRarity2026 4 0.492 1.000 0.385 1116.000
Md Ragib Amin Nihal Capped Rarity K-Center nihalRarity2026 5 0.481 0.900 0.128 1134.450
Jaqueline Garcia-Yi AFL garciaYiAFL2026 6 0.477 1.600 1.333 1031.800
Clea Parcerisas Coreset KMeans parcerisasQuantiles2026 7 0.469 1.000 4.095 1012.250
Baseline Coreset (Baseline) biodcase2026T4 8 0.464 1.000 5.838 1023.600
Clea Parcerisas Coreset Eigenvalues parcerisasQuantiles2026 9 0.435 0.800 7.800 1010.600
Baseline TypiClust (Baseline) biodcase2026T4 10 0.421 1.000 5.076 958.950
Baseline Margin Multilabel (Baseline) biodcase2026T4 11 0.408 1.000 0.002 1263.650
Baseline Random (Baseline) biodcase2026T4 12 0.401 1.000 0.001 966.050
Clea Parcerisas All Quantiles KMeans parcerisasQuantiles2026 13 0.397 1.000 0.330 992.050
Clea Parcerisas Balance Class by Clusters (Eigenvalues) parcerisasQuantiles2026 14 0.393 0.800 0.110 1039.000

Subsets

Team ranking across the four data subsets: ATBFL, HSN, POW, UHH. Showing the AULC (mAP).

Submission information Rank Subsets (AULC mAP)
Corresponding Author System name Technical
report		 Official
rank		 Rank
score		 ATBFL HSN POW UHH
Gabriel Dubus ADU-MMR dubusAdaptive2026 1 0.507 0.502 0.625 0.480 0.421
Hugo Magaldi CARE-DPP magaldiCARE2026 2 0.505 0.500 0.601 0.490 0.428
Haoyu Wang PB-MFS wangPBMFS2026 3 0.499 0.491 0.601 0.481 0.422
Wayne Yang Safe Rarity yangSafeRarity2026 4 0.492 0.492 0.589 0.475 0.412
Md Ragib Amin Nihal Capped Rarity K-Center nihalRarity2026 5 0.481 0.489 0.568 0.474 0.392
Jaqueline Garcia-Yi AFL garciaYiAFL2026 6 0.477 0.483 0.552 0.462 0.411
Clea Parcerisas Coreset KMeans parcerisasQuantiles2026 7 0.469 0.496 0.555 0.440 0.386
Baseline Coreset (Baseline) biodcase2026T4 8 0.464 0.477 0.531 0.453 0.394
Clea Parcerisas Coreset Eigenvalues parcerisasQuantiles2026 9 0.435 0.457 0.484 0.435 0.364
Baseline TypiClust (Baseline) biodcase2026T4 10 0.421 0.477 0.430 0.421 0.358
Baseline Margin Multilabel (Baseline) biodcase2026T4 11 0.408 0.457 0.438 0.419 0.319
Baseline Random (Baseline) biodcase2026T4 12 0.401 0.465 0.374 0.433 0.333
Clea Parcerisas All Quantiles KMeans parcerisasQuantiles2026 13 0.397 0.461 0.402 0.393 0.332
Clea Parcerisas Balance Class by Clusters (Eigenvalues) parcerisasQuantiles2026 14 0.393 0.447 0.389 0.408 0.328

Technical reports

Adaptive Diversity-Uncertainty Active Learning with Redundancy Control for Bioacoustic Event Classification

Dubus, Gabriel and Magaldi, Hugo and Gros-Martial, Anatole
Eco-Anthropologie, Mus\'eum National d'Histoire Naturelle, UMR7206, CNRS, Paris, France; Centre d'Etudes Biologiques de Chiz\'e, UMR 7372, CNRS, La Rochelle Universit\'e, France

Method summary
Embeddings PerchV2
Decision rule global-confidence adaptive uncertainty-diversity weighting + greedy Maximum Marginal Relevance batch selection

Adaptive Frontloaded Learning

Garcia-Yi, Jaqueline
Ecosonus

Method summary
Features swarm-optimised (PSO-style) diversity/coverage/density warm start
Embeddings PerchV2
Decision rule swarm-based warm start + margin-based sampling with a front-loaded, budget-aware acquisition schedule

Determinantal Point Process Sampling for Bioacoustic Active Learning

Magaldi, Hugo and Dubus, Gabriel
Eco-Anthropologie, Mus\'eum National d'Histoire Naturelle, UMR7206, CNRS, Paris, France

Method summary
Embeddings Perch v2
Decision rule annealed uncertainty-novelty weighting + DPP-based batch diversification with candidate-pool exploration and adaptive acquisition batch sizing

A Density-Gated Coverage Sampler for Cross-Domain Active Learning in Bioacoustics

Nihal, Ragib Amin and Yen, Benjamin and Ashizawa, Takeshi and Nakadai, Kazuhiro
Institute of Science Tokyo, Department of Systems and Control Engineering, Tokyo, Japan

Method summary
Features capped class-deficit rarity; budget-ramped mean-margin uncertainty (kNN label surrogate); metadata-only location stratification
Embeddings Perch v2 (cached PCA-128 projection)
Decision rule density-gated greedy k-center (farthest-first) coverage, gated by average labels-per-sample

Are Whales Too Big to Fly?

Bordoux, Valentin and Cuyx, Bram and Parcerisas, Clea and Schall, Elena
Wageningen University \& Research (WUR); Flanders Marine Institute (VLIZ); Alfred Wegener Institute (AWI)

Method summary
Features quantile-based class-confidence sampling; class-balancing by cluster/confidence; coreset (k-means / eigenvalue warm-up variants)
Embeddings Perch 2.0
Decision rule feature-space coverage / diversity sampling (best: scikit-learn coreset with k-means or eigenvector-based warm-up)

PB-MFS: Multi-Funnel Selection with Pareto-Balanced Uncertainty and Diversity for Bioacoustic Active Learning

Wang, Haoyu and Dou, Han and Chen, Hui and Du, Yushan and Tu, Liushi and Pan, Ningning and Zhang, Hao and Wu, Jingyao and Li, Gang and Huang, Gongping
Southwestern University of Finance and Economics, China; MiLM Plus, Xiaomi Inc.; IASP Lab, Wuhan University, China; Massachusetts Institute of Technology, USA

Method summary
Features multi-funnel candidate filtering (CF, MM, LCB, NC, LCI criteria); PCA-based farthest-first cold start
Embeddings perch_v2
Decision rule Pareto-balanced uncertainty-diversity greedy batch selection over multi-funnel survivor set

Safe Rarity-Aware k-Center Active Learning for Bioacoustic Classification

Yang, Xiangwen (Wayne)
Monash University, Melbourne, Australia

Method summary
Features top-k binary entropy uncertainty; k-center diversity; predicted-prevalence rarity
Embeddings Perch v2
Decision rule adaptive gamma-gated uncertainty-diversity-rarity score mixing + greedy score-plus-diversity candidate pool selection