Passive acoustic monitoring (PAM) is a powerful tool for studying biodiversity and ecosystem health, but analyzing vast amounts of acoustic data remains a challenge. Traditional feature extraction methods often lack adaptability, are static and might fail to incorporate domain knowledge effectively. 

To address this, an interactive contrastive learning framework that integrates human-in-the-loop feedback to refine learned features or representations. In this thesis, Contrastive learning, which optimizes feature spaces by pulling similar sounds together and pushing dissimilar ones apart, is enhanced through user-guided similarity annotations. This interactive process mimics a similarity search system but with user involvement, iteratively improving embeddings for ecologically meaningful soundscape analysis. 

By bridging automated learning with expert input, our approach enhances interpretability, reduces dependence on extensive labeled datasets, and improves clustering, classification or other downstream tasks.  

Objective: 

Develop an interactive contrastive learning framework for passive acoustic monitoring (PAM) that refines soundscape representations through human-in-the-loop feedback and suitable augmentation methods.  

Design a user-in-the-loop contrastive learning system that integrates expert feedback to iteratively improve the quality of learned audio representations. 

Design an active learning strategy: The system should prioritize ambiguous or informative pairs where user feedback can provide the most significant improvement. 

Evaluate the impact of interactive learning on downstream tasks such as species classification and unsupervised clustering by comparing it against traditional representation learning methods.