题目：Diffusion Models Respond the Duty Call from Causal Discovery 

报告人：王如心 副研究员(中国科学院深圳先进技术研究院)  

摘要：Learning a faithful directed acyclic graph from observed data is a promising but challenging problem. Recent work formulate the problem as a continuous optimization problem which can be disassembled as two parts: solving an inverse problem and satisfying an acyclicity constraint. However, solving an inverse problem is vulnerable to instability. In this paper, we propose a novel minimization objective with a delicately designed variance-negotiation regularizer which utilizes the variance variable in additive noise models to accomplish regularization on DAG.  

With a series of evolution on the minimization objective for causal discovery, we find optimizing the objective analytically equals to training a Denoising Diffusion Probabilistic Model, albeit of some incompatibility problems. This finding inspires us to fix all incompatibility problems and coin a novel diffusion model named DAG-Invariant Denoising Diffusion Probabilistic Model (D^3PM), which play the roles of not only causal learner but also data generator. Extensive empirical experiments on synthetic and real datasets are conducted. D^3PM achieves record-breaking performance on all small-scale and large-scale benchmarks.  

报告人简介：王如心，中国科学院深圳先进技术研究院副研究员，中国科学院大学博士生导师，中国科学院特聘研究岗位（特聘骨干人才），深圳市“鹏城孔雀计划”特聘岗位，中国运筹学会数学与智能分会副秘书长，图论组合分会青年理事，中国人工智能产业发展联盟医学人工智能委员会工作组专家。主要研究方向包括模式识别，因果机器学习，图像处理，多模态表征计算等，在知名国际学术期刊、会议发表论文30余篇，获广东省科技进步二等奖、深圳市科技进步一等奖，深圳市优秀科技论文等奖励，作为负责人主持国家重点研发计划青年科学家项目、国家自然科学基金面上项目、青年项目、广东省自然科学基金面上项目，深圳市优秀科技创新人才培养项目以及华为横向课题等多项国家、省部级项目以及大型科技企业合作攻关课题。  

报告时间：2024年9月14日（周六）上午10:00-11:00  

报告地点：教二楼727

联系人：胡晓楠