Machine learning
The goal of the machine learning team is to characterise and compare whistle repertoires within and between dolphin species.
By using machine learning, we can explore the complexity of dolphin communication by identifying acoustic patterns hidden within large collections of dolphin recordings. Over the past year, our focus has been on supervised learning, developing and fine-tuning machine learning models that can classify dolphin species based on their whistle sounds using spectrogram images. Three machine learning models were explored: a convolutional neural network (CNN) built from scratch, a pre-trained ResNet18 model, and a vision transformer. These models were integrated into a streamlined pipeline for audio processing, spectrogram generation, and model training, testing, and optimisation (Figure 1).
Figure 1: A model of the machine learning pipeline, divided into data input, data processing, model training/testing, and hyperparameter tuning.
Moving forward, the team is shifting towards unsupervised learning to investigate variation within a single species. By clustering whistles into types without predefined labels, we aim to uncover natural groupings and examine how these may relate to behaviour, context, or environment. A starting point will be using unsupervised learning methods like Hierarchical DBSCAN, alongside transfer learning with previously trained models from the past year.