Machine learning to predict killer whale (Orcinus orca) behaviors using partially labeled vocalization data

Abstract

Orcinus orca (killer whales) exhibit complex calls. In a call, an orca typically varies the frequencies, varies the length, varies the temporal patterns, varies their volumes, and can use multiple frequencies simultaneously. Behavior data is hard to obtain because orcas live under water and travel quickly. Sound data is relatively easy to capture. This paper studies whether machine learning can predict behavior from vocalizations. Such prediction would help scientific research and have safety applications because one would like to predict behavior while only having to capture sound. A significant challenge in this process is lack of labeled data. This paper works with recent recordings of McMurdo Sound orcas where each recording is labeled with the behaviors observed during the recording. This yields a dataset where sound segments—continuous vocalizations that can be thought of as call sequences or more general structures—within the recordings are labeled with potentially superfluous behaviors. This is because in a given segment, an orca may not be exhibiting all of the behaviors that were observed during the recording from which the segment was taken. Despite that, with a careful combination of recent machine learning techniques, including a ResNet-34 convolutional neural network and a custom loss function designed for partially labeled learning, a 96.1% general behavior label classification accuracy on previously unheard segments is achieved. This is promising for future research on orca behavior as well as language and safety applications.