报告人：张琼
报告地点：数学与统计学院415教室
报告时间：2024年12月13日星期五13:30-14:30
报告摘要：
Mixture models are widely used for model-based clustering, and they have encountered new challenges in scaling inference methods to accommodate modern large-scale datasets. The Split-and-Conquer (SC) learning techniques are developed to effectively handle large datasets, and they have been studied extensively for models with Euclidean parameter spaces. Unlike regular models whose parameter spaces are Euclidean, the parameter space of finite mixtures is formed by discrete distributions with a fixed number of support points, which makes the conventional SC approaches infeasible. In this talk, we introduce novel SC approaches for the distributed learning of finite mixtures under regular distributed learning systems and when there are Byzantine failures in the system. Experiments are conducted on both simulated and real-world datasets to show the efficiency of the proposed method. Our proposed method also has better statistical and computational performance than some existing approaches.
主讲人简介：
张琼2015年本科毕业于中国科学技术大学少年班学院。2022年博士毕业于加拿大英属哥伦比亚大学统计系。2022年9月起加入中国人民大学统计与大数据研究院并担任助理教授。目前的研究兴趣包括：混合模型、分布式学习、联邦学习等。她的研究论文发表在Journal of Machine Learning Research, IEEE Transactions on Information Theory, ICCV等机器学习期刊和会议上，现主持国家自然科学基金青年项目。
