The team used a combination of pattern recognition and classification algorithms to analyze 8,719 recordings of codas from around 60 whales collected by the Dominican Sperm Whale Project between 2005 and 2018. They found that the way whales communicate is neither random nor simple, but structured according to the conversation context, allowing them to identify unique vocalizations that had not previously been noticed.
Rather than relying on more complex machine learning techniques, the researchers opted to use classical analytics to approach the existing database with a fresh perspective.
“We wanted to have a simpler model on which to base our hypotheses,” Sharma says.
“The nice thing about the statistical approach is that you don’t have to train a model, it’s not a black box. [the analyses are] “It’s gotten easier to do,” says Felix Effenberger, a senior AI research adviser at the Earth Species Project, a nonprofit that studies how to use AI to decipher nonhuman communication. But he notes that machine learning is a great way to speed up the process of discovering patterns in datasets, so employing such methods could be useful in the future.
Dan Chernoff/Project CETI
The algorithm transformed the clicks in the coda data into a new kind of data visualization the researchers call exchange plots, revealing that some codas contain extra clicks. These extra clicks, combined with variation in the duration of their calls, appear in exchanges between multiple whales, which the researchers say suggest that codas carry more information and have a more complex internal structure than previously thought.
“One way to think about our discovery is that the sperm whale communication system, which has previously been analyzed as being like Egyptian hieroglyphics, is actually more like writing,” says Jacob Andreas, an associate professor at CSAIL who worked on the project.
The team isn’t sure whether what they found can be interpreted as the equivalent of letters, tongue positions or sentences in human languages, but they are confident that there were many internal similarities between the codas they analyzed, he says.
“This tells us that there are many more types of codas, or distinctions between codas, that whales can clearly recognise.[and] The data shows that people were completely unaware of it.”
The team’s next step is to build a language model of whale calls and explore how those calls relate to different behaviors. They also plan to develop a more general system that can be used across species, Sharma says. Taking a communication system that we know absolutely nothing about, figuring out how it encodes and conveys information, and slowly beginning to understand what’s being communicated could have a variety of applications beyond whales. “I think we’re just starting to understand some of these things,” she says. “It’s early days, but we’re slowly making progress.”
Understanding what animals are saying to each other is a major motivation for such projects, but if you want to understand what whales are communicating, there’s a big hurdle: you need experiments to prove that such an endeavor would actually work, says Caroline Casey, a researcher at the University of California, Santa Cruz, who has studied elephant seal vocal communication for more than a decade.
“Since the advent of AI, there’s been a resurgence of interest in decoding animal signals,” Casey says, “It’s very hard to prove that a signal actually means the same thing to an animal as a human would. This paper does a very good job of illustrating the nuances of an animal’s acoustic structure, but going the extra mile to figure out what the signal means is very hard.”
