Publication
Active and Transfer Learning for Efficient Identification of Species in Multi-Label Bioacoustic Datasets
Hannes Kath; Thiago Gouvea; Daniel Sonntag
In: Proceedings of the 2024 International Conference on Information Technology for Social Good. ACM International Conference on Information Technology for Social Good (GoodIT-2024), September 4-6, Bremen, Germany, Pages 22-25, ISBN 9798400710940, ACM, 2024.
Abstract
The complex system of life on Earth, biodiversity, provides the essential resources for human survival. However, humanity is the primary driver of species extinction, accelerating the extinction rate to 100-1000 times higher than pre-industrial times. To combat this alarming trend, detailed information on biodiversity is needed, making effective monitoring technologies essential. Passive Acoustic Monitoring (PAM) has emerged as a key technology for scalable wildlife monitoring. While PAM is effectively used to record vast amounts of acoustic data, the automatic identification of species remains an unsolved challenge. This elaboration formally describes the problem of identifying as many species as possible in an unlabelled PAM dataset while examining the fewest samples. A pilot study is conducted to investigate the potential of four approaches combining transfer learning with adapted uncertainty or diversity active learning sampling strategies. The findings of this study indicate that uncertainty-based sampling strategies yield superior performance to random sampling. In contrast, the diversity-based strategies used demonstrate inferior performance, with improvements observed when fine-tuning the embedding space using already labelled data. This study lays the groundwork for future research aimed at iteratively fine-tuning the embedding space in combination with uncertainty and diversity methods.