Tracks in the Sand: AI Settles a Long-Running Paleontology Dispute

Paleontology has always resembled reading tea leaves, except that instead of a cup you have a multi-ton block of sandstone, and instead of leaves, a barely visible depression left by someone's paw a hundred million years ago. Remember how as kids we tried to guess the shapes of clouds? Scientists do much the same when they look at fossilized tracks, known as ichnites. The problem is that the same dinosaur leaves entirely different tracks depending on whether it ran through sticky clay or walked on dry sand. Add erosion, the pressure of overlying layers and the researcher's own subjective view, and you get a perfect recipe for scientific wars that drag on for decades.

A group of German researchers decided it was time to end the romance and bring in mathematical precision. They turned to deep learning to systematize, once and for all, what used to be a matter of intuition. The scientists singled out eight typical features, a kind of "digital passport" for a footprint, that includes the specific angles of the toe segments, the ratio of toe lengths and the depth of certain pressure zones. The human brain tends to look for familiar patterns where there are none, but a neural network is impartial. It sees geometry where we see just a hole in a rock.

Why does this matter right now? We have reached a point where the number of fossils found exceeds scientists' physical capacity to process them. Museum storerooms hold thousands of track slabs that were classified by eye back in the last century. AI makes a large-scale review of these collections possible. The very first tests showed that the algorithm can tell dinosaur species apart even when a footprint is only partially preserved. This opens the way to a global database in which every newly found track would be instantly compared against all known finds on the planet.

Track analysis gives us what bones cannot: dynamics. A skeleton tells us what an animal looked like when it died. A track tells us how it lived: how fast it moved, whether it hunted alone or in a pack, and how it interacted with its environment. Digitizing that evidence with AI turns paleontology from a descriptive discipline into an exact data science. We can now model the gait of ancient reptiles with a fidelity that used to be available only to Hollywood special-effects artists, but now strict mathematics stands behind it.

Skeptics may of course say that we are trusting machines to interpret our past. But the irony lies elsewhere: the algorithms we build to design self-driving cars and facial recognition systems turn out to be best at explaining our deepest past. We are using the technology of the future to finally understand who ruled this planet long before the first hint of human intelligence. In a couple of years it may turn out that half the exhibits in your favorite museums are labeled incorrectly, and we will have to rewrite the textbooks.

The bottom line: AI is turning paleontology into data science, removing the human factor from the identification process. Are we ready for neural networks to rewrite the history of dinosaurs?