SoylentNews

aristarchus [soylentnews.org] writes:

Just a data problem? From LiveScience [livescience.com].

What do those clicking sounds mean?

"Jimmy's in trouble, down by the mill"? I always get my Flipper and my Lassie mixed up. But on to serious questions.

Sperm whales are among the loudest living animals on the planet, producing creaking, knocking and staccato clicking sounds to communicate with other whales that are a few feet to even a few hundred miles away.

This symphony of patterned clicks, known as codas, might be sophisticated enough to qualify as a full-fledged language. But will humans ever understand what these cetaceans are saying?

The answer is maybe, but first researchers have to collect and analyze an unprecedented number of sperm whale communications, researchers told Live Science.

With brains six times larger than ours, sperm whales (Physeter macrocephalus) have intricate social structures and spend much of their time socializing and exchanging codas. These messages can be as brief as 10 seconds, or last over half an hour. In fact, "The complexity and duration of whale vocalizations suggest that they are at least in principle capable of exhibiting a more complex grammar" than other nonhuman animals, according to an April 2021 paper about sperm whales posted to the preprint server arXiv.org. [arxiv.org]

This paper, by a cross-disciplinary project known as CETI (Cetacean Translation Initiative), outlines a plan to decode sperm whale vocalizations, first by collecting recordings of sperm whales, and then by using machine learning to try to decode the sequences of clicks these fellow mammals use to communicate. CETI chose to study sperm whales over other whales because their clicks have an almost Morse code-like structure, which artificial intelligence (AI) might have an easier time analyzing.

Now pay attention, this is not some runaway collection of letters and numbers! CETI is not "SETI", even though both are searching for intelligence, and sort of come together in a particular Star Trek movie.

Pratyusha Sharma, a data science researcher for CETI and a doctoral candidate in the Computer Science and Artificial Intelligence Laboratory at MIT, told Live Science more about recent developments in artificial intelligence and language models, such as GPT-3, which uses deep learning to construct human-like text or stories on command, and last year took the AI community by storm. Scientists hope these same methods could be applied to the vocalizations of sperm whales, she said. The only problem: these methods have a voracious appetite for data.

The CETI project currently has recordings of about 100,000 sperm whale clicks, painstakingly gathered by marine biologists over many years, but the machine-learning algorithms might need somewhere in the vicinity of 4 billion. To bridge this gap, CETI is setting up numerous automated channels for collecting recordings from sperm whales. These include underwater microphones placed in waters frequented by sperm whales, microphones that can be dropped by eagle-eyed airborne drones as soon as they spot a pod of sperm whales congregating at the surface, and even robotic fish that can follow and listen to whales unobtrusively from a distance.

But even with all this data, will we be able to decipher it? Many of the machine-learning algorithms have found audio more difficult to analyze than text. For instance, it might be challenging to parse apart where one word begins and ends. As Sharma explained, "Suppose there's a word 'umbrella.' Is 'um' the word or is it 'umbrell' or is it 'umbrella'?" The barriers between spoken words are more ambiguous and less regular, and patterns may therefore require more data to suss out.

And even with the data, the AI model could be pro-human biased, and is "suss" even a word? Maybe it is time we find Dory.

Original Submission