Research
Alongside my home lab, I work with the Imageomics Institute, an NSF HDR institute on extracting biological information from images with knowledge-guided machine learning. That work ties evolutionary morphology, comparative methods, and representation learning together so image-based traits are more useful for macroevolutionary questions.
Neural networks can learn characters from complex observations and support inference under models without tractable likelihoods, but they learn statistical regularities rather than the entities and processes biologists use to explain diversity. That mismatch is a design problem: architectures, losses, data, and optimization can impose biologically motivated constraints so learned representations stay commensurable with evolutionary theory.
I call networks designed and evaluated this way Deep Comparative Methods. Related applied work includes PhyloNN, which uses restructured autoencoders and multi-task learning to incorporate phylogeny and ontogeny into phenotypic models.
To evaluate evolutionary machine learning with known ground truth, I build synthetic image and mesh datasets from defined morphospaces and phylogenies. TraitBlender, a Blender add-on I developed, renders standardized museum-style specimens in bulk for benchmarking computer vision models under controlled evolutionary processes.
Meaningful comparative models need assumptions that match biological reality. Ontologies—knowledge graphs of hierarchical trait relationships—let us encode that structure directly in analyses. I have contributed to this through RPhenoscate and related Phenoscape/SCATE work on computable anatomical traits.