On blisteringly hot desert sands, researchers crawled on their hands and knees avoiding fist-size cacti littering the ground. Their goal: collecting bones and teeth of some of the earliest known primates to shed light on the adaptations at the root of the evolutionary lineage that led to humans. The fossils, though, are the size of a fingernail or smaller, and they are scattered over an area of about 10,000 square kilometers in the rocky desert of Wyoming’s Great Divide Basin.
That’s a lot of ground to cover, especially on all fours and in searing heat. So the scientists are relying on a tool never tried before in paleontology: artificial intelligence. Such an approach might be able to pinpoint fossil troves in their giant needle-in-a-haystack quest and suggest new strategies for fossil hunting. It then remained for them to wander to the middle of the desert to see if their innovation led them on a wild goose chase or not.
Normally, discovering fossils depends largely on luck. Paleontologists can take educated guesses as to where to search—trekking down dry stream beds to look for bones that might have eroded off slopes, for instance—but they mostly depend on walking around to see what catches the eye. If they are lucky, they can cover ground in bucking and bouncing jeeps down dirt roads set up by oil and gas companies. In any case, traditional approaches can be challenging, lengthy—and fruitless.
Increasingly, paleontologists are relying on technology to narrow their search for fossils. For instance, Google Earth has helped identify sites in South Africa containing fossils of the ancient hominid Australopithecus sediba.
But instead of inspecting satellite imagery by eye for potential sites, paleontologist Robert Anemone and remote-sensing specialist Jay Emerson of Western Michigan University and their colleagues have developed a way to automate the operation using an artificial neural network, a computer system that imitates how the brain learns. Their aim was to take advantage of how brains, both natural and artificial, quickly learn and recognize patterns, such as what fossils look like.
Training the artificial brain
Artificial neurons are components of computer programs that mimic real neurons in that each neuron can send, receive and process information. Researchers first train the networks by feeding data to the artificial neurons and letting them know when their computations solve a given problem, such as reading handwriting or recognizing speech. The networks then alter the patterns of connections among these neurons to change the way they communicate with one another and work together. With such practice, the networks figure out which arrangements among neurons are best at computing desired answers.
The neural network presented the promise of locating fossil-rich sites “without walking over miles and miles of barren outcrops,” says paleontologist John Fleagle of Stony Brook University. “It could save lots and lots of time and expense in the field.”
That’s why Anemone and his colleagues were out in the Wyoming desert with a neural network running on a laptop computer. It analyzed visible- and infrared-light satellite and aerial images of the Great Divide Basin, which included 100 known fossil sites. They first let the network know that 75 of these areas were fossil-rich so it could learn what this kind of site looked like. When they had it search for the other 25 sites, it correctly spotted 20 of them, raising hopes that it could identify new candidates.
Filling gaps in the primate record
The researchers were hunting fossils dating to the late Paleocene and early Eocene epochs, about 55 million to 50 million years ago, when the Rocky Mountains were first rising and the climate was significantly warmer and wetter on average than today. Back then a large freshwater lake dominated the dig region, with streams flowing to it from the surrounding mountains. The area was home to crocodiles, turtles, lizards, fish and lots of mammals, including very primitive rodents, horses and bats as well as primates similar to modern lemurs, tarsiers, lorises and galagos.
Today, the area is mostly dry sagebrush scarred and pocked with gullies, buttes and dunes. Pronghorn antelope run alongside cars and groups of elk occasionally dash in front of them. Roaming stallions greet campers in the morning with thunderous snorts, and falcons occasionally dive at the visitors to keep them away from nests. The area seems mercifully free of venomous snakes, but thunderstorms can destroy tents and clog trails with slippery mud that can trap a truck.
Scientists have worked in this desert since 1994 and have unearthed roughly 10,000 mammalian fossils from the 100 localities now known. Even so, “this basin has been a blank mark on the map for a long time when it comes to ancient primates,” remarks paleontologist Brett Nachman of the University of Texas at Austin. Many more mammalian fossils are needed.
Anemone’s neural network pointed out several places to search. Initially, these proved fruitless—the scientists unearthed many fossils at the first recommended sites, but not the kind they wanted. The researchers had the neural network search for fossils in areas that past geologic surveys declared were in the Wasatch Formation—former lakeshore and riverside areas where they expected to unearth primate fossils. But on arrival at the first dozen or so sites, it was clear the original surveys were in error. Instead, those locations were actually in the Green River Formation—former lakebed areas with many aquatic fossils but few mammal bones.
On the next-to-last day of field work this past summer, however, the researchers looked at three more suggested sites, ones they were sure were located in the Wasatch Formation. After hiking for about an hour across the seemingly flat desert, they came across boulder-strewn hollows where they uncovered a cache of mammalian fossils, including teeth from the extinct five-toed horse Hyracotherium.
“This is the first successful test of an artificial neural network to find fossils,” Anemone says. “It led us right through the sagebrush.”
Intriguingly, these three sites were depressions in the soil that were all but invisible from a distance on the ground, unlike the hills the researchers normally inspect for fossils. “We would never have found these sites without the neural network—they’re so out of the way and hidden,” Anemone says. “I was worried about whether the neural network would work,” he adds. “It sure felt good pulling those fossils out.”
Beyond the black box
Fleagle cautioned that neural networks are essentially black boxes—one could never be sure how the systems arrive at their findings. “It would be good to know what features the algorithm is actually using to identify the fossil areas,” he says. “Are they the same features that a geologist could also identify from a map or in the field?”
As a result, scientists see the neural networks only as stepping-stones to a more analytical approach to fossil hunting. In fact, Anemone and his colleagues are now directly scanning how fossil sites appear in visible and infrared light in the hope they can predict for themselves how these locations might differ from other areas. Besides the Great Divide Basin, they hope to conduct such research in Africa, possibly helping to look for fossils of monkeys or even early hominids.
Although the team will not be relying on the neural network, Anemone feels it was invaluable at proving that computer models can find fossils. Their odyssey guided by artificial intelligence showed it was possible to detect fossils remotely—now it is time for human intelligence to lead the way in creating a new and better way to predict where fossils are hidden. “The future of paleontology lies in this technology,” Anemone says.