John Thickstun is an assistant professor of computer science in the Cornell Ann S. Bowers College of Computing and Information Science.
Before arriving at Cornell, Thickstun was a postdoctoral scholar in the Stanford Artificial Intelligence Laboratory at Stanford University. He received his Ph.D. in computer science and engineering from the University of Washington.
What is your academic focus?
My focus is on machine learning, and more specifically, on generative models.
Could you describe your research?
I work to improve generative technologies across a wide range of modalities: text, images, music, and audio. I am particularly interested in methods that control or regulate the behavior of these models. I think about control from both the perspective of a user, who hopes to use a model to accomplish a task, and from the perspective of a model provider or policymaker, who hopes to broadly regulate the behavior of a model. I am also interested in applications of generative models that push beyond the standard text and image modalities, including music technologies.
What inspired you to pursue a career in this field?
Machine learning asks me to think about problems from so many different perspectives. There are computational and engineering challenges to be solved. There is algorithmic work informed by mathematical insights. And there are increasingly pressing socio-technical questions about deployments of this technology. It is exciting to be able to think about such a diverse set of questions!
What courses are you most looking forward to teaching?
I'm enthusiastic to co-teach introductory machine learning this fall and give students their first exposure to the field. Looking forward, I hope to develop new courses on the theory of generative models, and on music technologies.
What scientific questions are you looking to answer next or areas you plan to explore?
I'm actively working to develop new music technologies powered by generative models. I'm also revisiting work I did a few years ago on source separation (the "cocktail party problem") and thinking about how to improve these methods using state of the art diffusion models. More abstractly, I'd like to better understand the interplay between discrete and continuous models and data, using this understanding to inform algorithmic innovations.