Identifying variation in dinosaur footprints and classifying problematic specimens via unbiased unsupervised machine...

Area of Science:

• Paleontology
• Machine Learning
• Evolutionary Biology

Background:

• Traditional fossil classification relies on supervised methods, which can introduce researcher bias.
• Machine learning offers potential for objective fossil identification and resolving paleontological debates.

Purpose of the Study:

• To apply an unsupervised machine learning technique to classify dinosaur footprints.
• To objectively identify trackmakers and resolve debates on the origins of early birds and ornithopods.

Main Methods:

• Utilized a disentangled variational autoencoder network on a database of 1,974 dinosaur footprints.
• Identified eight key shape variation features differentiating tracks.
• A posteriori labeled tracks and analyzed morphospace using distance metrics.

Main Results:

• Achieved 80-93% agreement with expert identifications after unsupervised analysis.
• Classified Late Triassic-Early Jurassic bird-like tracks with modern and fossil birds.
• Grouped Middle Jurassic three-toed tracks with ornithopods, suggesting earlier origins.

Conclusions:

• Unsupervised machine learning provides a robust, less biased method for fossil track analysis.
• Findings support an earlier evolutionary origin for birds and ornithopods than previously indicated by body fossils.
• The DinoTracker app and source code facilitate broader application of this technique.