A child’s dinosaur obsession led to an AI with a very old mystery to solve

Gregor Hartmann’s route into palaeontology began with his young son Julius. According to Helmholtz-Zentrum Berlin’s account, the child’s fascination with dinosaurs led the physicist to read Stephen Brusatte’s The Rise and Fall of the Dinosaurs. Hartmann contacted its author, and a collaboration grew with Brusatte and footprint specialists Paige dePolo and Tone Blakesley.[2]

Hartmann normally applies machine learning to problems including electron motion and X-ray pulses. Dinosaur tracks presented a different kind of signal: an animal was gone, but a shape remained. The team built DinoTracker to examine what those shapes have in common.[2]

Their January PNAS paper reinforces an intriguing observation: disputed ancient tracks resemble bird footprints. Who made them remains unresolved.[1]

A footprint is an event

A foot enters ground, bears weight, moves and leaves. The resulting mark records that encounter. The same anatomy could leave different outlines in firm sand and soft mud; different feet could leave deceptively similar ones. Reading a footprint therefore means working backwards through an event, with some of its conditions missing.

Edinburgh’s account puts two sets of disputed prints at the centre of the story: marks from the Isle of Skye dating to about 170 million years ago, and birdlike tracks made more than 200 million years ago. An abundance of three-toed marks does not automatically resolve the identity of the animals crossing those landscapes.[3]

There is also a circularity problem. If a training collection labels every disputed print “bird”, a model can become proficient at repeating that decision. Its apparent agreement would add little evidence to the original dispute. A useful new comparison should have some opportunity to surprise its creators.

DinoTracker learned from 1,974 two-dimensional silhouettes without animal-group labels. Compressing and reconstructing shapes yielded eight interpretable dimensions, including toe spread, heel position and contact area, after rotation and mirroring were excluded. Researchers attached published identifications afterwards.[1]

That order of operations changes the question. Instead of asking a machine to reproduce a catalogue, the researchers first ask which shapes belong near one another. They can then compare the resulting neighbourhoods with the catalogue. Agreement is reassuring; disagreement gives them something specific to investigate.

The bird resemblance survives

For previously assigned tracks, agreement was about 80% when separating non-avian theropods from herbivorous ornithopods, and more than 93% when separating non-avian theropods from birds. “Non-avian” distinguishes birds’ other theropod relatives from the birds themselves. These are agreements with expert classifications, not independently established accuracies. Most of seven disputed early birdlike tracks fell in bird-dominated regions.[1]

The authors retain three explanations: very early birds, other dinosaurs with similar feet or movement, and ground conditions producing a birdlike trace. Their method uses outlines, leaving depth, size, substrate and age outside the model. Resolving the bird-origin mystery requires body fossils.[1]

That is still progress. “Looks like a bird” becomes a comparison readers can inspect: which aspects of the outline create the resemblance, and which other footprints occupy the same region? A disagreement with a familiar label can become a measurable pattern instead of a contest between two impressions.

The project’s public repository includes the footprint data, trained model and a Python application. Its interface lets users import a silhouette, inspect its encoded features, alter feature values and compare nearby examples. The authors provide installation instructions; this is downloadable research software, and the results reported here are the team’s, rather than a new test of the application.[4]

Moving a feature and watching a reconstructed footprint change also makes an abstract model easier to question. A reader can ask whether the distinction being emphasised has a plausible relationship to anatomy or could reflect the ground. The explanatory value lies in seeing what a comparison depends on.

Julius’s interest helped bring a physicist into a fossil argument. The software that followed leaves that argument open in a more useful form. Somewhere between the outline and its missing owner is evidence a new discovery could still change.

Report an error or suggest a correction