Adit Radha

I am interested in (1) advancing the mathematical foundations of AI and (2) developing novel, theoretically-grounded algorithms for biomedical applications.

Advancing mathematical foundations of AI: I am most interested in understanding feature learning (the ability of AI models to discover patterns from data). My current research is focused on discovering general mechanisms for feature learning (like the Average Gradient Outer Product – AGOP) that can help us better understand the structure in data that is learned by large-scale AI systems. Ultimately, the goal of this research program is to use these new feature learning mechanisms to develop a new class of interpretable, effective, and computationally-efficient AI models that can uncover structure in massive heterogeneous datasets.

Other AI / ML topics of interest: kernel machines, unsupervised / self-supervised learning, infinite width / depth limits of neural networks

Applications in biology: I work on developing algorithms to discover and characterize cellular programs across various disease contexts. My current research focuses on understanding these programs using million-to-billion scale single-cell RNA sequencing datasets. Ultimately, the goal is to leverage various genotypic and phenotypic readouts to understand how programs drive cell state and function and how these programs are altered in disease contexts. The outcome of this research program is to identify targeted therapies for disease based on disrupted cellular programs.