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

Machine learning holds great promise for classifying and identifying fossils, and has recently been marshaled to identify trackmakers of dinosaur footprints and address long-standing debates over whether some dinosaur tracks are the oldest birds or ornithopods (duck-billed herbivores and kin) in the fossil record, or alternatively were made by nonavian theropods. Existing methods in paleontology, however, require supervision and a priori labeling of training data by researchers, which can lead to bias. We employ an unsupervised machine learning technique for recognizing inherent patterns in shape data, using a disentangled variational autoencoder network, to a database of 1,974 footprints, spanning a diversity of dinosaurs across their evolutionary history, including modern birds. Our neural network identified eight features of shape variation that most differentiate these tracks: overall load and shape (amount of ground contact area), digit spread, digit attachment, heel load, digit and heel emphasis, loading position, heel position, and left-right load. With the unsupervised process finished, we a posteriori labeled each track based on published expert judgments, plotted them into morphospace, and applied distance metrics to group means and nearest neighbors, which showed 80 to 93% agreement with expert identifications. Controversial Late Triassic-Early Jurassic bird-like tracks group with fossil and modern birds and some Middle Jurassic three-toed tracks with ornithopods, supporting an older origin for these groups than recorded by body fossils. We provide an app, DinoTracker, to make this process accessible, and source code that can be adapted to other cases where paleontologists or biologists are studying patterns of shape variation.