Elephants Under Attack Have An Unlikely Ally: Artificial Intelligence
A few years ago, Paul Allen, the co-founder of Microsoft, published the results of something called the Great Elephant Census, which counted all the savanna elephants in Africa. What it found rocked the conservation world: In the seven years between 2007 and 2014, Africa’s savanna elephant population decreased by about a third and was on track to disappear completely from some African countries in as few as 10 years.
To reverse that trend, researchers landed on a technology that is rewriting the rules for everything from our household appliances to our cars: artificial intelligence. AI’s ability to find patterns in enormous volumes of information is demystifying not just elephant behavior but human behavior — specifically poacher behavior — too.
“AI can process huge amounts of information to tell us where the elephants are, how many there are,” said Cornell University researcher Peter Wrege. “And ideally tell us what they are doing.”
There are two kinds of elephants in Africa: savanna elephants, which were counted by Allen’s census, and forest elephants, which the census couldn’t account for because that elephant lives beneath a thick rainforest canopy. Even at the level of the jungle, Wrege says, losing a forest elephant is easy to do. “Sometimes you see them, let’s say, 15 meters [16 yards] away from you and then they move 5 meters into the forest and you can’t see them,” he said. “Somehow they just disappear.”
Researchers at Cornell University have been studying the forest elephant for years, trying to figure out — like Allen did with the savanna elephant — how many there are and how fast they are being killed. Given how stealthy the forest elephants are, Wrege began to think that rather than look for them, maybe he should try something a little different: Maybe he should listen for them.
To do this, Wrege had 50 custom audio recorders made. He divided the rainforest into a grid and headed to Central Africa. His team hung the custom recorders every 23 feet to 30 feet in the treetops, just a little higher than an elephant could reach with its trunk while standing on its hind legs. And then they hit record. Three months later, they would return to the forest, locate the recorders, change the batteries, put in new audio cards, and start all over again.
As the months wore on, the recorders were collecting hundreds of thousands of hours of jungle sounds, more than any team of graduate students could realistically listen to — which meant Wrege had another problem: How could he sort through all these recordings to find the elephant voices he wanted?
“In AI circles this is known as the ‘cocktail party problem,’ ” said computer scientist Josephine Wolff, who is now a professor at the Fletcher School at Tufts University. “At a party with a lot of background noise, the human brain can focus on a specific person’s voice and amplify that above all the other voices. AI can do the same thing.”
In fact, there’s a subset of AI — something called a neural network — that is very good at this. A neural network is essentially a group of algorithms, or mathematical equations, working together to cluster and classify information and find patterns humans wouldn’t necessarily see. It is particularly good at working with images, so Wrege ran the audio through a software program that turned the recordings he had collected into spectrograms — ghostly little pictures of sound waves. He then had a company in Santa Cruz, Calif., called Conservation Metrics build him a neural network that could sort through the cacophony of jungle sounds and find elephants.
“Basically each of these ‘neurons’ in the network is determining how likely one piece of the spectrogram is to belong to an elephant call,” Wolff explains. “Neurons in the first layer have just the spectrogram to consider, and so will likely recognize things like pitch and other things it sees as defining characteristics of elephant calls.
