BADGE-Greedy-DPP: new active learning method boosts rare bioacoustic call detection

Bioacoustic call-type classification typically requires costly expert annotation of long audio recordings. To minimize this burden, active learning algorithms aim to select the most informative segments for labeling. The challenge is extreme sparsity and a long-tailed call distribution—most segments contain no target call, and rare calls are few yet critical. Existing batch selectors like k-means++ or MCMC-based DPP sampling offer no strong optimality guarantees.

In a new paper, Shiqi Zhang and colleagues introduce BADGE-Greedy-DPP, a deterministic batch selector that greedily adds segments whose BADGE gradient embeddings maximally enlarge the volume spanned by the selected batch. Because the log-volume objective is submodular, this greedy rule guarantees a batch value at least (1-1/e) of the optimum—a guarantee missing from prior BADGE heuristics. Additionally, the authors exploit the natural frame-level behavior of BADGE: prediction residuals weight the aggregated pseudo-gradient, so confidently-predicted no-call frames contribute little, while a single uncertain rare-call frame can set the segment’s direction. This elegantly resolves the temporal granularity mismatch between segments and informative frames. In 10 runs on a sparse, imbalanced hyena call-type dataset, BADGE-Greedy-DPP achieved the best overall and rare-call-type performance, outperforming MFFT and two vanilla BADGE traversals.