Angular Contrastive Loss for Few-Shot Learning in Bioacoustics Sound Event Detection

Abstract

The acceleration of global climate change has increasingly degraded the natural ecosystem and accelerated the loss of biodiversity year after year. Improvements in Machine Learning (ML) models and techniques has made it possible to develop reliable and effective wildlife monitoring tools in higher resolution. One such tool is Sound Event Detection (SED), which has the potential to help ecologists to locate and identify species in a certain area in real time. However, in a field where both human resources and annotated training data are often scarce, traditional supervised ML models often fail to address the need for reliable and effective detection tools that can work within this limitation. The Few-Shot Learning (FSL) framework is often used as a solution to train supervised ML models when high-quality annotated training data are scarce while adaptability of the model to unseen classes without significant performance regression is desired. The performance of a model trained with FSL relies heavily on the quality of feature embedding discrimination during training, which is why the choice of loss function in training phase is crucial to the performance of the model in downstream tasks. In this thesis, two loss functions designed specifically to optimise embedding space discrimination--Angular Contrastive Loss (ACL) and Supervised Contrastive Loss (SCL)--were explored and their performance were compared in an experiment. Experiment result shows that ACL outperforms SCL in the DCASE 2024 Development Set dataset, where the model trained using ACL achieved an F1-Score of 54.98% compared to 49.47% of baseline SCL. It was also found that even with lower finetuning epochs, ACL still outperforms SCL with 51.13%.