Much of the excitement about AI over the past decade has come from the achievements of neural networks — systems built on an analogy of how the human brain processes information through collections of neurons. Deep learning, in which data pass through many layers of a neural network, was what led to the creation of the chatbot ChatGPT. The sperm whale, elephant and marmoset studies, however, used earlier forms of AI known as decision trees and random forests.
A decision tree is a classification algorithm that looks like a flow chart. It might ask, for example, whether the sound it has been given has a frequency above a certain value. If yes, it might then ask whether the call lasts for a certain length of time, and so on, until it has decided whether the call matches the acoustic variables it was trained to look for using human-labelled data sets. A random forest is a collection of many decision trees, each constructed from a randomly chosen subset of the data.
Kurt Fristrup, an evolutionary biologist at Colorado State University who wrote the random-forest algorithm for the elephant project, says that tree-based algorithms have several advantages for this kind of work. For one, they can work with less information than is required to train a neural network, and even thousands of hours’ of recordings of animal calls is still a relatively small data set. Furthermore, because of the way that tree-based algorithms break down the variables, they’re not likely to be thrown off by mislabelled or unlabelled data.
The random forest also provides a way to verify that similar calls match: different calls that show the same features should each end up in the same ‘leaf’ of an individual tree. “Since there were on the order of a thousand of these trees, you get a fairly fine-grained measure of how similar two calls are by how often they landed in the same leaf,” Fristrup says.
An elephant reacts to the playback of a call that was originally addressed to her. Credit: Mickey Pardo
An elephant reacts to the playback of a call that was originally addressed to her. Credit: Mickey Pardo
It is also easier to see how a random-forest algorithm came to a particular conclusion than it is with deep learning, which can produce answers that leave scientists scratching their heads about how the model reached its decision. “Deep-learning models make it possible or even easy to get all kinds of results that we can’t really get any other way,” Fristrup says. But if scientists don’t understand the reasoning behind it, they might not learn “what we would have learnt had we got into it by the older, less efficient, and less computationally intense path” of a random forest, he says.
Despite this, the ability of a neural network to generalize from a relatively small, labelled data set and discover patterns by examining large amounts of unlabelled data is appealing to many researchers.
Machine-learning specialist Olivier Pietquin is the AI research director at the Earth Species Project, an international team headquartered in Berkeley, California, that is using AI to decode the communications of animal species. He wants to take advantage of neural networks’ ability to generalize from one data set to another by training models using not only a large range of sounds from different animals, but also other acoustic data, including human speech and music.
The hope is that the computer might derive some basic underlying features of sound before building on that understanding to recognize features in animal vocalizations specifically. This is the same way in which an image-recognition algorithm trained on pictures of human faces learns some basic characteristics of pixels that describe first an oval and then an eye. The algorithm can then take those basics and recognize the face of a cat, even if human faces make up most of its training data.
Olivier Pietquin (back row, second left) and other members of the Earth Species Project are attempting to decode animal communication. Credit: Earth Species Project
Olivier Pietquin (back row, second left) and other members of the Earth Species Project are attempting to decode animal communication. Credit: Earth Species Project
“We could imagine using speech data and hope that it will transfer to any other animal that has a vocal tract and vocal cords,” Pietquin says. The whistle made by a flute, for example, might be similar enough to a bird whistle that the computer could make inferences from it.
A model trained in this way could be useful for identifying what sounds convey information and which ones are just noise. To work out what the calls might mean, however, still requires a person to observe the animal’s behaviour and add labels to what the computer has identified. Identifying speech, which is what researchers are currently trying to achieve, is just a first step towards comprehending it. “Understanding is really a tough step,” Pietquin says.
