Ilyass Moummad Thesis defense

This thesis focuses on developing robust and transferable representation learning techniques for few-shot bioacoustic event detection and classification, addressing core challenges in deep learning such as domain generalization, domain adaptation, data scarcity, and class imbalance. Through the exploration of self-supervised invariant representation learning, we demonstrate that domain-agnostic data augmentations can yield informative and discriminative representations. A key focus of this work is the use of supervised contrastive learning to enhance model generalization across different species and acoustic environments.
Furthermore, we propose a novel supervised contrastive loss, inspired by prototypical networks, that reduces the computational complexity of the traditional supervised contrastive loss while maintaining