Plenary Talk: What Statistics and AI Offer Each Other?
This plenary presentation, delivered by Emmanuel Candès at ICASSP 2026, addresses the bidirectional synergy between statistical theory and artificial intelligence to construct more reliable, safe, and efficient AI systems. The primary goal is to examine how rigorous statistical frameworks—particularly conformal prediction and controlled inference—can be integrated into machine learning workflows to ensure high-quality, trustworthy outputs when automated models substitute for or guide real-world data collection.
The talk surveys key computational and statistical challenges across several real-world vignettes, including constrained training of large language models, data exhaustion, and AI-driven clinical trials. Central evidence highlights techniques for quantifying uncertainty in model predictions, verifying whether AI automated tasks correctly, and using machine learning predictions as effective surrogates or guidance mechanisms for costly data acquisition. The presentation demonstrates how statistical principles prevent spurious findings and guarantee rigorous error rate control in complex AI-driven decision pipelines.
Ultimately, the key takeaway is that statistical rigor is essential for turning black-box AI predictions into dependable engineering tools. By establishing theoretical guarantees around machine learning outputs, researchers and practitioners can deploy safer, more robust AI systems across critical scientific, biomedical, and industrial applications.
