The grammar induction approach I developed in my PhD AI EDAM was motivated by design, but the underlying problem — recovering structure from noisy examples of complex artifacts — is not specific to things humans make. Working with Phil LeDuc and Jonathan Cagan at CMU, we extended the method to biological systems, where rules are implicit, numerous, and very noisy FASEB’17.

More recently we applied the same machinery to medical imaging. With collaborators at CMU and the University of Pittsburgh, we induced grammars that describe vascular structure in brain angiograms and used them to classify anomalies JESMDT’22. The underlying method is covered by US Patent 11,899,669.

What I like about this line of work is that the same representational move — treating a complex artifact as the output of an unknown grammar — keeps paying off across very different domains. The structure is usually there; the question is whether we can let the data tell us what it is.