Research Project Summary:

The project aims at using deep generative models to generate images of planktonic foraminifera species. It will explore whether we can generate a latent space with meaningful distances in generative adversarial network settings relating to species. For example, can we generate novel images that a domain expert can confidently confirm to belong to the species generated? Could we interpolate in this space to generate images in-between two species where a domain expert would struggle to categorize? we would like to explore whether it would be possible to build a phylogenetic tree in this latent space. For example, would it be possible to learn what differentiates two species and how would this explanation compare to that of the domain expert?

Planktonic Foraminifera identification is crucial for many research areas. However, only a handful of resources exist to train new students in the difficult task of discerning amongst closely related species, resulting in diverging taxonomic schools that differ in species concepts and boundaries. We propose to use supervised deep learning methods for automatic classification and generation of Forams in order to learn what phenotypic features differentiate them from a machine learning perspective. We will explore whether we can generate a latent space with meaningful distances relating to species. Finally, it would be an interesting task to see in what ways this ML perspective is similar and different from a human perspective.