Reproduce “Aggregation Strategies for Efficient Annotation of Bioacoustic Sound Events Using Active Learning” on OpenResearch

About this paper

Richard Lindholm, Oscar Marklund, Olof Mogren, John Martinsson

Lindholm et al. introduce Top K Entropy, an active learning aggregation strategy for efficient annotation of bioacoustic sound events. This method, applied to a segment-based detection model using YAMNet features, achieved model performance comparable to full dataset training with a 92% reduction in annotation effort, outperforming other baseline strategies.

Key insights

• Prioritizing files based on segments with the highest uncertainty peaks (Top K Entropy) is more effective for Active Learning in bioacoustics than using average uncertainty across all segments.
• Active Learning, particularly Top K Entropy, provides greater annotation efficiency gains in challenging conditions such as low Signal-to-Noise Ratios and sparse sound event distributions.
• The performance of Top K Entropy is robust to the specific choice of K and inherently selects diverse batches, reducing the need for additional diversification techniques.

Summary by alphaXiv.