What We Do
During the project, tasks were split into 4 sub-teams:
- Biology
- Database design
- Computer Science
- Outreach
1. Biology
The main roles of those interested in the biology aspect of the project were to select, extract, and analyse dolphin whistles. This involved using RavenLite, PAMGuard, and ARTwarp
(refer to ‘Methods’ section for more information on each software and how they’re used).
Ultimately, analysis of whistles in this way will aid the conservation of dolphins, and allow researchers to learn more about the geographic differences in dolphin whistle types. See the ‘Home’ page for more information on our aims and goals.
2. Database design
Those involved in organising the datasets containing the dolphin whistles involved:
- Improving the folder framework in the VIP OneDrive folder.
- Ensure that the data is easily accessible and easy to navigate, allowing new VIP members to transition into the project with ease.
Currently, new software is being developed (called MetaEmbed), which uses metadata to help locate original data and eliminate the need for confusing and error-prone file names.
3. Computer Science
There are currently multiple exciting tasks undertaken by the Computer Science sub-team. These include:
CetaceXplain
CetaceXplain is a Python module used to understand how machine learning models classify cetacean whistle contours. Machine learning models are computer based algorithms used to recognise patterns. In this case, the patterns were dolphin whistles contours (Figure 1). Recognising patterns in these contours allows machine learning models to classify these whistles.
CetaceXplain uses the SHAP Python library. SHAP uses Shapley values, derived from economic game theory that is now being widely applied to machine learning and artificial intelligence (AI). SHAP, in this module, indicates pixels on spectrograms that are considered important to a species classifier’s decision making process.
Figure 1: Spectrograms highlighting the pixels that are important for a machine learning (ML) model. Each of the grey columns represent the model’s classification of the image for different dolphin species. The red pixels indicate a positive influence on the machine learning model and the blue pixels indicate a negative influence. The left image uses the model trained on coloured images, whereas the right image uses greyscale images. Removing colour may improve the model by learning that colour choice for displaying spectrograms is irrelevant.
Transfer Learning
Transfer Learning involves using machine learning models that have already been trained on previous data. These models are used to classify new dolphin whistle datasets.
On this VIP, the convolutional neural network (CNN)-based Xception architecture was used, which was pre-trained to classify many images into many different classes. In particular, the architecture was modified and trained to distinguish between whistles produced by bottlenose, common, and melon-headed dolphins. Current results are promising, with an accuracy of ~90%, but this is currently being improved further.
ARTwarp
ARTwarp is a MATLAB-based program that uses an ART2 neural network to categorize tonal animal sounds. More information on how and why we use ARTwarp, as well as improvements being made to it, can be found in the ‘Methods’ section.
4. Outreach
In this area of the project, students focused mainly on designing and creating content for this website. In addition, citizen science communication was explored via reading and discussion of relevant literature.