My google scholar will usually be the most up-to-date way to see what I’ve been working on.

The leitmotif of my research is the combination of ideas from machine learning, physics, and information theory to help understand complex systems like human behavior and biology.

Most recently, this has led to an interest in diffusion models, like DALL-E-2 (see an example image generated by DALL-E-2). These methods give us unprecedented power to model complex probability distributions, and to generate realistic samples from these distributions. The mathematical foundation of diffusion models is based on non-equilibrium dynamics, and we showed in a recent preprint that it also has deep ties in information theory. The information-theoretic formulation of diffusion gives us many easy implications which we are exploring in ongoing work, including improved mutual information estimation, supervised learning, ways to accelerate sampling, and more.

For other work on latent factor discovery, energy-based models, graph representation learning, theory of generalization in learning, information estimation, and applications to neuroimaging and gene expression, please see my recent papers.