Science1 publisher2 min readPublished
Algorithm trained on Amboseli rumbles predicts elephant behavior better than chance
Researchers including ElephantVoices' Mickey Pardo used machine learning on hundreds of Amboseli rumble recordings to predict elephant behavior above chance. Rumbles made in similar situations seem to share acoustic properties, a measurable start on cataloguing elephant calls.
The Scientist · Science desk
What happened
- The analysis centered on six kinds of rumble: coos to calves, distant contact calls, let's-go calls before moving, cadenced calls at rest or meals, and two separate greetings.
- Many recordings caught several elephants rumbling at once, producing overlapping calls the researchers refer to as choruses.
- Within choruses, rumbles sometimes grew progressively more alike, and Keen said this convergence appeared only when the elephants seemed to be coordinating group action.
- In February 2020 Pardo watched a Samburu mother named Matisse rumble to her newborn, but the study's recordings came from a separate population in Amboseli.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- constraint Until per-class accuracy is published, the result describes how rumbles differ on average and cannot yet be used to label any single new recording with confidence.
- cost Extending the catalogue to new groups or sites means paying observers to log behavior alongside each recording, because the model learns from those field labels.
- capability If whole-chorus analysis transfers, researchers could study group calling in orcas and crows, both named by Keen as candidates, without separating each caller first.
Each recording was paired with what the elephants were observed doing, and the algorithm had to work out the behavior from the sound alone [4][10]. If a let's-go rumble before a move and a cadenced rumble at a meal sounded the same, its guesses would land at chance. They came out above chance, though the predictions were not entirely accurate [4].
Above chance is a minimum. How far above it matters, and that depends on how many behaviors were in play and how often each turned up. The analysis named six kinds of rumble [1]. If the model was choosing among that many situations, it could clear chance comfortably and still mislabel most of the calls it heard. The coverage does not report the model's accuracy, the exact count of recordings, or whether the algorithm sorted calls into types on its own or was given the categories.
The category names suggest the categories were given, since each one describes a situation [3]. I think the study, as reported, supports a narrower claim than decoding. The situation an elephant is in leaves a trace in its rumble, and a model can pick that trace up better than guessing. Scientific American reports that what these calls mean has long been a mystery [14], so a measurable acoustic signal is progress even at this resolution.
Keen, a co-author and senior research scientist at Earth Species Project, said: "It is pretty exciting that we're able to train a model to look at an entire chorus and have some interpretation of exactly what's going on" [6]. The evidence is thinnest on vocal convergence, a pattern also seen in parrots and dolphins [7]. The study links matched calls with apparent coordination [8]. Whether the matching helps a group decide or only happens while it moves is an open question, and Keen said the idea that it signals agreement needs further testing [9].
Arik Kershenbaum, a University of Cambridge zoologist who was not involved, said: "This study emphasizes how vital it is to pair new computational analyses with real behavioral observations in the wild" [10]. "Animals communicate to influence each other's behavior, and so you can't just study the sounds in isolation," he said [11].
What to watch
- Per-class accuracy in the Scientific Reports paper, including whether the model can tell the two greeting rumbles apart.
- Follow-up work testing whether elephants that match their rumbles then act together more often than those that do not.
- Whether a model trained on Amboseli recordings holds up on another population, such as the Samburu elephants Pardo observed.