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

arXiv cs.AI · 18 hours ago

Researchers developed BADGE-Greedy-DPP, a batch selection method for active learning in bioacoustic call classification that greedily maximizes the volume of gradient embeddings to identify the most informative segments for expert annotation. The method guarantees batch quality of at least (1-1/e) fraction of optimal value for the log-volume objective and addresses temporal granularity mismatches by weighting frames based on prediction confidence. Testing on a sparse, imbalanced hyena call-type dataset across 10 runs, BADGE-Greedy-DPP outperformed baseline strategies including MFFT and vanilla BADGE variants on both overall and rare-call-type classification performance.