A Two-Way Street: How Statistical Thinking Powers AI Efficiency and How AI Inspires New Statistical Inference?
Abstract
We have entered an era where deep learning and foundation models are transforming data analysis, increasingly handling prediction tasks that were traditionally the domain of statistical modelling. This rapid shift raises a fundamental question: How should statistics evolve in a landscape dominated by large-scale AI? In this talk, I argue that rather than becoming obsolete, traditional statistical principles are essential for overcoming the natural limits of brute-force scaling. I present a research program driven by a dual perspective: applying statistical thinking to solve engineering bottlenecks in modern AI, and conversely, leveraging AI paradigms to inspire new statistical methodologies. I will illustrate this synergy through three chapters of my research:
- Statistical Efficiency for AI Systems: I first demonstrate how Mixture Reduction grounded in optimal transport addresses computational redundancy, enabling the compression of 3D computer graphics models by 90% while preserving geometric fidelity. I further apply this rigor to Federated Learning, resolving label switching and utilizing Empirical Likelihood to transform central servers into "intelligent routers" that leverage, rather than suppress, data heterogeneity.
- AI Inspires New Statistics: Turning the direction of influence, I explore how In-Context Learning (ICL) redefines statistical inference. We show that foundation models trained via ICL can outperform specialized statistical methods in a wide range of tasks.
Biography:
Qiong Zhang is currently an assistant professor at the Institute of Statistics and Big Data at Renmin University of China. She did her PhD in Statistics at the University of British Columbia. Her work bridges statistics and artificial intelligence, with a focus on developing methods that make data a more powerful storyteller. Currently, her research explores:
- Mixture reduction with applications to federated learning, computer graphics
- Empirical likelihood for intelligent federated learning, handling label noise
- Tabular foundation models and in-context learning for statistical tasks