LULU GARCIA-NAVARRO, HOST:
A few years ago, researchers fanned out across Africa to count all the elephants. But one species presented a challenge, the forest elephant whose population has been threatened by poachers. As part of a special series about the technologies that watch us, NPR's Dina Temple-Raston looks at how conservationists are using something called neural network in hopes of saving the elephants.
GARCIA-NAVARRO: The forest elephant lives in central African rainforests so dense, the enormous animals are incredibly hard to spot. So Peter Wrege, an elephant researcher at Cornell University, decided to try something new to just listen for them.
PETER WREGE: If we know how often an elephant gives a vocalization, and we can record it, then we can spread recorders over a big area and record their vocalizations and use those numbers to count them.
DINA TEMPLE-RASTON, BYLINE: Wrege had 50 custom audio recorders made and divided the forest into a grid.
WREGE: We put recorders seven to 10 meters up in a tree hanging from a tree limb.
TEMPLE-RASTON: Thirty feet isn't an arbitrary height. That happens to be just a little higher than an elephant can reach with his trunk while standing on his hind legs. So Wrege's team strapped these audio recorders in place every 5 kilometers in the forest and then just hit record.
(SOUNDBITE OF RAINFOREST NOISE)
TEMPLE-RASTON: This is the actual audio from one of those recorders.
WREGE: We record anything that makes an acoustic signature, including things that are not vocalizations like the pounding on a tree buttress by chimpanzees.
(SOUNDBITE OF CHIMPANZEES SCREAMING)
TEMPLE-RASTON: So these tape recorders can record for about three months, after which Wrege sends the teams back into the forest to climb the trees and...
(SOUNDBITE OF ARCHIVED RECORDING)
UNIDENTIFIED PERSON: Ayo.
WREGE: ...Bring the recording units back down, change batteries...
(SOUNDBITE OF ARCHIVED RECORDING)
UNIDENTIFIED PERSON: Ayo.
WREGE: ...Change the SD card so we have the recordings and put them back up in the trees.
(SOUNDBITE OF ARCHIVED RECORDING)
WREGE: Yo.
TEMPLE-RASTON: Remember, his recorders have been running 24 hours a day for three months. And that's a lot of audio to get through, a lot of chimps and frogs and birds, just to hear what he really wanted, the rumbling voices of forest elephants.
(SOUNDBITE OF ELEPHANT RUMBLING)
WREGE: Just like, whoa - we've got it.
(SOUNDBITE OF ELEPHANT RUMBLING)
TEMPLE-RASTON: So Wrege's giant acoustic census captured those elephant sounds just like he'd hoped. But then he started thinking, how can I possibly get through everything I've recorded? He had a classic big-data problem. At the same time, a behavioral ecologist in California named Matthew McKown was having a big-data problem of his own. He wasn't an elephant researcher. His specialty was birds. And he had recorded huge amounts of bird songs.
MATTHEW MCKOWN: You know, sitting on mountaintops with tape recorders - he was recording the songs of reasonably obscure birds.
(SOUNDBITE OF BIRDS CHIRPING)
MCKOWN: I was one of the few people in the world, I think, that used mini-discs. I was super excited about mini-disc recorders.
TEMPLE-RASTON: And he recorded all the bird songs and then thought to himself, just like Wrege, now what? So he started to look for ways that he could process all that data. That's when he and a colleague found a subset of artificial intelligence called neural networks. Neural networks are called neural networks because they resemble the interconnected structures of the brain. The easiest way to think about it is in terms of layers one on top of another. They cluster and classify things. And they're best at doing that when those things are in the form of pictures.
So McKown and Wrege turned their recordings into something called a spectrogram, a little ghostly picture of the soundwave. And then they fed those pictures into a neural network to see if it could find birds and elephants. The neural network learns to identify and count them in a step-by-step way. Neurons in the first layer of the network would likely recognize something simple, like pitch or modulation, something that might characterize a bird or an elephant or...
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: ...For the purposes of our discussion, a particular instrument...
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: ...Say violins.
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: And defined violins in an orchestral piece of music - it might have to start to look for that modulation. The next layer of neurons might look for a range of notes that it's learned is associated with violins.
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: And it thinks, I recognize some notes that a violin might play. And then one of these neurons lights up. Then middle layers would build on that, focusing in on the other qualities that might be associated with a violin, like the presence of strings. It might pick out a cello or a piano for their string-like sound.
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: There can be hundreds and hundreds of layers each getting progressively more refined, some, perhaps, filtering out things that the network decides definitely are not violins like, say, percussion.
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: It says to itself, this is not a violin.
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: And all this gets melded together with one layer after another, taking the network closer and closer to its goal - recognizing the violin - until, at the end, the network looks at all the information that it used to filter down to the violin. And it makes a statistical calculation. What's the likelihood that this pattern I've identified is a violin or an elephant or a bird? - 50%? - 80%? - until it gets what it wants...
(SOUNDBITE OF MUSIC)
TEMPLE-RASTON: Violins.
(SOUNDBITE OF MUSIC)
MCKOWN: This whole field is called deep learning.
TEMPLE-RASTON: That's Matthew McKown again. And he started a company called Conservation Metrics, which builds these kind of neural networks for conservationists.
MCKOWN: And, specifically, what we do is turn that information into actionable information for people on the ground.
(SOUNDBITE OF ELEPHANTS RUMBLING)
MCKOWN: That actually is a fantastic example of two females who are performing what we call a greeting ceremony.
TEMPLE-RASTON: These are the kinds of sounds a neural network would find in those jungle recordings. In this case, this is a sequence of rumbles that happens when elephants who haven't seen each other in a while come together again.
(SOUNDBITE OF ELEPHANTS RUMBLING)
MCKOWN: I think it's very much like if you run into a friend on the street, you know, that you haven't seen for a while. Whoa. How are you? And, oh, I'm OK. What about you? Oh, it's not so good. I've lost my job. Oh, my God. You know, who knows what they're really saying? But it's that - perhaps that kind of thing.
WREGE: Remember, Wrege was doing all of this so he could count the forest elephants. And that count is still going on. The neural network they built is still training on those recordings. And Wrege has found that getting an accurate count depends on lots of things - weather, the season, where in the forest they're listening. So there isn't a precise number yet. But the Elephant Listening Project has already started to learn things it didn't know. For example, Wrege's team has found that elephants don't go into some parts of the forest during specific times of year. That means anti-poaching teams don't have to go there because poachers aren't going to find elephants there anyway. Those are the kinds of helpful things Wrege expects to discover in the years ahead.
Do you think that AI is going to save the elephant?
WREGE: I actually do. It definitely is going to be our salvation.
TEMPLE-RASTON: AI and elephants - the beginning of a beautiful friendship - Dina Temple-Raston, NPR News. Transcript provided by NPR, Copyright NPR.
