What is a species? A new machine learning approach bridging taxonomic concepts in biology and palaeontology

Despite over 250 years of species descriptions since Linnaeus, the challenge of accurately delimiting species remains unresolved – particularly between fossil and living taxa, where data types and availability differ greatly. This has resulted in multiple competing species concepts and complicated biodiversity comparisons across time.

I present preliminary results from a project developing a novel machine learning framework for species delimitation that integrates image data from extant species – where boundaries are molecularly established – with related fossil species. Our study focuses on the renowned Plio-Pleistocene Viviparus beds of paleo-lake Slavonia (SE Europe), which preserve an iconic viviparid species flock. Since the 19th century, over 40 morphospecies have been described from these deposits, many serving as key biostratigraphic markers. The striking morphological diversification spans from small, smooth-shelled forms to large, highly ornamented shells. As a baseline for morphological diversity, we use a modern viviparid species flock from long-lived lakes Yunnan (China), exhibiting a similar range of variation.

To delineate species, we train a Siamese Convolutional Neural Network, which can learn from relatively small, labelled datasets, generalize to previously unseen classes, and perform classification with minimal retraining. A specific aim is to assess how varying degrees of fossil preservation affect model performance, offering potential for broader paleontological applications. Ultimately, we seek to reconstruct biodiversity patterns across geological time, infer diversification dynamics within the Slavonian viviparids, and critically evaluate the biostratigraphic utility of traditional morphospecies. Our novel system should be readily transferable towards a variety of species groups.