Learning a new language is hard, especially when no human speaks it. That’s the challenge facing an international team of researchers trying to decode sperm whale communication.
The Cetacean Translation Initiative (CETI) is a five-year, multidisciplinary project that will deploy networks of underwater microphones, drones, and even robotic fish to track sperm whales and record their vocalizations. Meanwhile, on land, researchers including Berkeley computer scientist Shafi Goldwasser, M.S. ’81, Ph.D. ’84, and Berkeley linguist Gasper Begus will apply machine learning to try to understand what the whales are saying.
As Goldwasser, a Turing Award winner who directs the Simons Institute for the Theory of Computing at Cal, explained to Berkeley News, it won’t be an easy task. “For CETI, we need to significantly extend the theory and practice of unsupervised language translation, where no correct translation examples are given, to a setting where our prior knowledge on what the whales may be communicating about is limited, and we can’t run controlled experiments.”
Sperm whales possess the largest brains of any animal on Earth and communicate through a variety of sounds including ultra-loud clicks produced in patterns called codas. Researchers believe these codas are learned, like human language, not innate, and that different social groupings of whales have different dialects.
Begus, director of the Berkeley Speech and Computation Lab, has developed artificial intelligence models designed to learn human speech much the way children do: unsupervised, by imitation and intuition. He is testing whether similar models could decipher whale communication.
He also hopes this work will convince people of the importance of preserving the whales’ habitat, which has been severely impacted by human activity. “If we get to know sperm whales better by learning their communication and the full scope of their cognitive and social life, it’s harder for us as a species to treat them like nonsentient beings and destroy them.”