主题：Agent Statistics: A Novel Statistical Theory for Trustworthy AI
主讲人：严晓东 西安交通大学
主持人：姜云卢 暨南大学
时间：2025年5月15日（周四）上午11:00-12:00
地点：暨南大学石牌校区经济学院大楼（中惠楼）323室
摘要
Task-driven artificial intelligence technology differs fundamentally from traditional data-driven or model-driven statistical methodologies. It achieves specific goals by perceiving the environment, making decisions, and taking actions, with a distinct problem-solving approach. However, there is currently a lack of a comprehensive statistical framework and theory in the field of task-driven machine learning. This report will introduce the mathematical model of task-driven statistical learning, which leads to the discovery and proof of several paradoxes: Independent + Independent = Not Independent, Good + Bad = Better, and Normal + Normal = Non-Normal. The scientific significance of these paradoxes lies in the limitation of traditional data-driven statistical methodologies in achieving task-driven goals. Therefore, we have pioneered Strategy Limit Theory to explain the statistical laws under the optimal decision sequence in task-driven machine learning, and named this direction Agent Statistics or Task-Driven Statistics. Additionally, the report will discuss the mathematical theoretical logic of Strategy Limit Theory and the advantages of its statistical results, starting from two scientific problems: statistical estimation and statistical testing. Through application scenarios from Didi and Huawei, it will highlight the unique advantages of models established under task-driven statistical thinking.
主讲人简介
严晓东，西安交通大学数学与统计学院教授，博士生导师，入选国家级青年人才项目和校内青拔A类支持计划，滴滴盖亚学者, 研究方向为智能体统计学，包括智能计算和智能推断等，目前兼任中关村软联智能算法委员会秘书长，学术成果发表在统计学著名期刊JRSSB,AOS,JASA和经济学著名期刊JOE等。在高等教育出版社以独立主编出版了《机器学习》《数据科学实践基础-基于R》和《大模型学习科研手册》三部教材或专著。
欢迎感兴趣的师生参加！
校对| 姜云卢
责编| 彭毅
初审| 姜云卢
终审发布| 何凌云
（来源：暨南大学经济学院微信公众号）
