Uncertainty-Aware Human Preference Study and Machine Learning Predictions.

Seminar in Mathematics

Abstract: Recent advances in machine learning(ML)/artificial intelligence (AI) have broadly impacted fields from social science to healthcare, economics, etc. Firstly, technologies used by AI systems rely heavily on aligning with human preferences through ranking algorithms. On the other hand, many of their downstream applications often operate as black-box processes. This dependence raises two critical questions about its reliability: How can we ensure that ranking results genuinely capture underlying signals rather than noise? How can we address the inherent opacity of ML/AI models and their potential for generating hallucinations---outputs that seem plausible but are factually incorrect?

In this talk, I will explore these challenges from a statistical perspective. First, I will discuss methods for quantifying uncertainty in ranking algorithm outputs by examining individual-level rankings, top-choice sets, and preference variations across different communities. Second, I will present two novel approaches to mitigate unreliable ML/AI predictions: (1) constructing reliable prediction sets for ML predictions that can accommodate diverse data structures, including images and social networks, and (2) devising methods to select robust uncertainty measures for evaluating Large Language Model outputs.