Statistics in the Age of AI: Theory, Methods, and Data

Abstract

Artificial Intelligence (AI) has surged in popularity, creating both opportunities and challenges for statistics. In this talk, I will present three recent directions from my lab that reflect our efforts to engage with the age of AI. First, I will discuss theoretical results for generative models, including statistical foundations connecting latent dimension, approximation error, and model complexity; robustness to data contamination during recursive training; and the effects of synthetic data on downstream tasks. Second, I will discuss a method that uses embeddings from large language models to enhance high-dimensional hypothesis testing, motivated by problems in cancer genomics where traditional methods are underpowered. I will also discuss extensions to genetic studies, for which we curated annotations for 8.9 billion genetic variants from the human genome, and obtained embeddings of these 8.9 billion variants for downstream tasks. Finally, I will switch to an infrastructural view, introducing STimage-1K4M, one of the first and largest publicly available spatial transcriptomics datasets curated by my group, consisting of 1,149 slides and more than 4 million pathology image–gene expression pairs across 50 tissue types. This resource has been downloaded over 290,000 times on Hugging Face and has facilitated the training of multiple foundation models.