Data Collection and Statistical Inference in the Age of Data & AI
Recent advances in AI offer exciting opportunities to reduce the cost of data collection. Large language models (LLMs), for example, are increasingly used as scalable stand-ins for human judgment in tasks like model evaluation and survey research. Yet, the imperfections of AI—ranging from hallucinated facts in generative models to fragility in predictive systems—pose challenges, especially when downstream decisions and discoveries depend on AI outputs. In this talk, I will discuss approaches to leverage AI’s potential while preserving scientific rigor.
In the first part, I will present prediction-powered and active inference: methods for valid statistical inference assisted by AI predictions. In the second part, I will introduce probably approximately correct (PAC) labels, a paradigm for labeling datasets more efficiently using AI. Together, these methods illustrate how AI can accelerate the scientific process without compromising reliability. I will demonstrate their versatility through applications involving state-of-the-art LLMs, AlphaFold, and more.
Tijana Zrnic is a Ram and Vijay Shriram Postdoctoral Fellow at Stanford University, affiliated with Stanford Data Science and the Department of Statistics. Tijana obtained her PhD in Electrical Engineering and Computer Sciences at UC Berkeley and a BEng in Electrical and Computer Engineering at the University of Novi Sad in Serbia. Her research establishes foundations to ensure data-driven technologies have a positive impact; she has worked on topics such as AI-assisted statistical inference, performative prediction, and mitigating selection bias.
Paper: https://www.science.org/stoken/author-tokens/ST-1540/full
About the Seminar Series
The last few years have seen a substantial increase in the reported success of machine learning (ML), and generative artificial intelligence (AI). These impact practices in delivering services from financial institutions to entertainment and medicine. However, scientific research also increasingly relies on large data sets, whose analysis leverages ML/AI. This seminar series aims to investigate if and how the paradigm for scientific research has changed or should change to incorporate these new tools and the possibilities they open.
A diverse group of scholars engaged in scientific research, method development, and historical and epistemological investigations will give a 50-minute presentation, followed by discussion.
The event is open to all. Stanford students and postdocs have the opportunity to engage more directly with speakers and topics by enrolling in the Canvas course here.
