Training AI for Social Good

Songs of Adaptation, a project of Future Generations University, has several different facets. Those of you who have been following us see content about birds, bioacoustics, climate change, community engagement, and the mention of Artificial Intelligence or AI. The term AI may conjure images of robot apocalypse or computers who win Jeopardy. In contrast, the AI model we are training is more a well-tuned ear.

When we gain data from our recording stations, we run it through a basic bioacoustics analysis software that can cluster similar sounds together. It puts together everything that sounds alike, including similar-sounding bird calls. This first machine is not the most “intelligent.” There are numerous things that confuse this program, such as rain and wind. So, a person on our team steps in and listens to the clusters made by the first machine. This is a process called data labeling.

A Songs of Adaptation team member, typically a local expert, tries to identify what the sounds are and adds a label to the audio clip. A label might include “rain”, “wind”, or be a specific species name. This can be difficult. Multiple birds could be calling at the same time, the bird can be almost too faint to hear, or there are many birds in the data that sound nearly identical. However difficult, our team has to sift through and make determinations as data labeling creates the material we need to train our AI model.

We feed the labeled clusters to the AI model, so it begins to learn to associate the sounds it hears to the labels we assigned. This process is called training the model.  We are still fine-tuning this process because there are a lot of factors to account for and sometimes the machine “hears” differently than we do. It may confuse similar sounds or be unable to identify when two sounds are mixed.

This process is iterative. Our team and experts are checking and re-checking the accuracy of our labels and the accuracy of the model to ensure the results are useful. As of right now, the AI model has been partially trained on three different species. It will be exciting to see how it works and to soon be able to let the machine do a large part of the analysis for us. As we continue gathering data, improving our process, and involving local people, we will be able to better engage and support communities in their growing resilience to a changing world. We are moving down a path of community-informed, data-driven adaptation to climate change.

Stay tuned for more in-depth posts about the technical component of our work.