Skip to main content

Generative models have the potential to transform society, from reducing costs of digital content creation to democratizing access by lowering barriers to entry for creative work. However, the poor usability of even the most promising models poses difficulties in their safe deployment in real-world applications. In this talk, I will discuss my research on controllable generation to address this challenge. In contrast to existing approaches which try to steer models via interfaces such as natural language, I propose to directly extract intangible user preferences from a set of examples using classical density ratio estimation techniques. I will show how to estimate such likelihood ratios accurately, by drawing inspiration from diffusion models and bridging the gap between the two distributions via an infinite continuum of intermediate bridge densities. Applying these techniques leads to improvements in bias mitigation for AI safety, where a variant of this approach has been deployed in pre-training DALL·E 2. I will conclude with a discussion of my work on personalized music generation, as well as interesting avenues for future work in societal applications.

Bio: I am a final year Ph.D. candidate in Computer Science at Stanford University advised by Stefano Ermon, where I’m affiliated with the SAIL and StatML groups. My research is centered around machine learning with limited labeled supervision, and is currently focused on developing techniques for better adaptation and controllability in deep generative models. I ground my methodological work in societal applications motivated by problems in sustainability and fairness.

My research is supported by the NSF GRFP, Stanford Graduate Fellowship, the Qualcomm Innovation Fellowship, and the Two Sigma Diversity PhD Fellowship. I completed my undergraduate studies in CS-Stats at Columbia, where I worked on problems in computational biology as part of the Pe’er lab.

I previously interned at Google Brain in 2019 as part of the Magenta project. In my free time I’m an avid tennis player, runner, and food enthusiast!