Robust Weakly Supervised Bird Species Detection via Peak Aggregation and PIE

A contemporary challenge in sound event detection is developing robust models with weak supervision. However, the problems arising from weak annotations are severe, especially in fields like bioacoustics, where the datasets contain non-negligible missing and incorrect labels, among other issues. Current methods developed on short and curated audio benchmarks misrepresent the difficulties of dealing with such scenarios. They leverage the Multiple-Instance-Learning framework, which relies on pooling functions for the many-to-one mapping of instances to their shared label. Unfortunately, non-standard assumptions under this framework have led to unreliable recipes that mix many sources of influence, resembling a murky semi-supervised approach with ill-posed objectives that do not address the fundamental problems of weak supervision in challenging and long recordings. This work introduces a novel weakly supervised training framework using peak-finding and attention aggregation. The significance of this straightforward and robust approach is demonstrated on a challenging bioacoustics dataset for bird species classification. Also, an effective data augmentation technique naturally stems from this framework to further enhance the models. The core idea of this method is to increase the fidelity of the input signals regarding their weakly assigned labels while keeping the sources of influence separate for incremental improvement. Finally, a reliable strategy via posterior analysis is suggested for truthful model comparison, which shows some metrics might mislead us on noisy and weakly labeled datasets. The result shows a clear improvement for such tasks and establishes a common practice for future weakly supervised audio models.