Scrolling through social media, you may have spotted content from a faux Keanu Reeves to Jim Carey channeling Jack Torrance or Nicholas Cage as everyone from Yoda and Dr. No, to the T-100 Robot from Terminator 2: Judgement Day.
Hyper-realistic videos and images like these — digitally altered photos, video or audio also called deepfakes — are notorious for social media pranking, but the technology has serious scientific applications, too, including ongoing efforts to study and conserve endangered species like the North Atlantic Right Whale.
Dave Johnston is the director of the Marine Robotics and Remote Sensing Lab at Duke University’s Nicholas School of the Environment, and he and his team collect a lot of data using satellite imagery and drones. He said AI technology is an efficient tool to break down that information once AI is trained on the specific task asked of it.
“You need a lot of examples to be able to teach it what to look for, and so for a lot of animals where we don't have lots of imagery to be able to train that model, we have to turn to other techniques to be able to help us and one of the ways is to use deep fakes, diffusion models, other types of models to create synthetic imagery that is hyper realistic and would allow us to train those models to make them better at detecting these rare species," he explained.
Traditionally, researchers had to use their own eyes to scour satellite and aerial images to track the species they were following, but now Johnston said AI detection tools can speed up the process.
There are only about 400 North Atlantic Right Whales left, after commercial whaling in the 1800s brought the species to the brink of extinction, and they never fully recovered. Humans continue to be the biggest factor, with many whales killed by boat strikes and fishing gear entanglement. Because their habitat covers waters from Canada to New England, and waters down the East Coast to Florida, Johnston says there’s a lot of ocean to scour for a relatively small number of the large creatures.
“We worked on a project with folks up in Canada and on the East Coast in the US here to try and use some of these machine learning models to find right whales in satellite imagery, very high-resolution satellite imagery," he said, "And Right whales are rare and the satellite takes pictures of the ocean that are giant, hundreds and hundreds of square kilometers, and it's kind of like trying to find a needle in a haystack.”
While Johnston’s research is focused on the marine environment, he said the technique can also be applied to research of other rare species as well – like the red wolf, which is only found in the wild in eastern North Carolina, with just about 30 of the critically endangered animals living in the wild in the Alligator River and Pocosin Lakes National Wildlife Refuges.
"If we are developing systems to be able to try and detect them, it may be possible to create deep fake models, deep fake images, to be able to train a better model for camera traps, say, to be able to detect those animals," Johnston said, "You could apply the same thing to other rare species that might be found in places where you just don't have the ability to collect the data that you need to train the model.”
But Johnston said the AI models are producing predictions rather than definitive scientific findings. “When we run one of these models, it predicts that there is a whale in that photo, say a right whale, in a satellite image and it will give us some sort of confidence level of that, and those don't really become detections until we decide, right?" he said, "So, we're the arbiters of the truth.”
He said it’s not far different from the way casual users employ Adobe Firefly or Dall-E to generate AI images. Johnston said, “You can put in a text prompt. You can say ‘Give me a picture of a whale in Antarctica’ and it will do a pretty good job of presenting you a realistic image of a whale near an iceberg or something, but it doesn't have enough fine scale detail to be really, truly accurate.”
An AI hallucination is when an artificial intelligence system generates information that is incorrect or false, but presents it confidently as factual, and Johnston said, “Sometimes these systems will actually hallucinate images of whales that are really weird. You know, you'll have whales with the flippers pointing the wrong way, or a whale with a tail on both ends instead of a head and a tail on one end. Those kinds of things. And, so, it's interesting to see the range of output that the models produce and it also reminds us that we need to be really careful as we use this tool, to make sure that it's actually working the way we think it is.”
Johnston added that because data centers that power A-I are energy- and water-intensive, they are using generative A-I conservatively and in targeted ways in order to lessen the impact on the environment. He also said the intersection between computer science and environmental sciences is only going to grow in the future.
