Improving data labeling efficiency for deep learning-facilitated bioacoustics monitoring

Over the last decade, deep learning has proven invaluable for classifying data with complex spatial and temporal relationships. Unfortunately, its utility in the context of bioacoustics monitoring has been limited by the unavailability of large, labeled datasets of species vocalizations. To explore solutions to this problem, various deep learning architectures and techniques were evaluated for their ability to reduce the data labeling efforts required to characterize the distinct sound stimuli present in two different acoustic environments. Located around Lake George, NY, these sites were acoustically monitored nearly continuously for 12 months. Commonly employed techniques such as transfer and semi-supervised learning were then analyzed for their ability to reduce the amount of labeled data necessary to achieve state-of-the-art classification results. Meanwhile, cross-corpus training was used to provide automatic “pre-labels” for these datasets, reducing the amount of total time associated with data labeling efforts. A hierarchical neural network was also implemented in order to reduce the performance costs associated with encountering sound stimuli in the test dataset that was not captured in the training dataset, perhaps as a result of the reduction techniques outlined.