哈尔滨工业大学（深圳）学术讲座

演讲人Speaker：     崔恒建（CUI Hengjian）

题目Title: A Class Sensitivity Feature Guided T-type Generative Model for Noisy Label Classification

时间Date：2025年 3月15 日       Time：9:30 am – 11:00 am

地点Venue：线上#腾讯会议：211-941-465

可线下参会： H 栋 301室

内容摘要Abstract： 

Large-scale datasets inevitably contain noisy labels, which induces weak performance of deep neural networks (DNNs). Current methods hardly focus on different extents of distortion of different features in latent space, which compromises validity in label noise scenarios. To solve this problem, we analyze characteristic distortion extents of different high-dimensional features, achieving the conclusion that features vary in their degree of deformation in their correlations with respect to categorical variables. Aforementioned disturbances on features not only reduce sensitivity and contribution of latent features to classification, but also bring obstacles into generating decision boundaries. To mitigate these issues, we propose class sensitivity feature extractor (CSFE) and T-type generative classifier (TGC). Based on the weighted Mahalanobis distance between conditional and unconditional cumulative distribution function after variance-stabilizing transformation, CSFE realizes high quality feature extraction through evaluating class-wise discrimination ability and sensitivity to classification. TGC introduces student-t estimator to clustering analysis in latent space, which is more robust in generating decision boundaries while maintaining equivalent efficiency. To alleviate the cost of retraining a whole DNN, we propose an ensemble model to simultaneously generate robust decision boundaries and train the DNN with the improved CSFE named SoftCSFE. Extensive experiments on three datasets, which are the RML2016.10a dataset, UCR Time Series Classification Archive dataset and a real-world dataset Clothing1M, show advantages of our methods.

个人简介(About the speaker):

崔恒建，现为首都师范大学教授，博士生导师。中国科学院系统科学研究所博士毕业。在大数据统计建模、高维统计及其稳健统计理论和方法、统计机器学习、金融统计、以及质量管理等领域取得过许多重要的研究成果，发表论文180余篇，其中包括发表在国际顶级的统计和计量经济学杂志JASA、 AoS、JRSS(B)、Biometrika和JoE上。主持国家自然科学基金重点项目、主要参加教育部重大科研基金项目、科技部等项目。现担任《数学学报》和《应用数学学报》中、英文版以及《Statistical Theory and Related Fields》编委，中国现场统计研究会副理事长，全国工业统计教育研究会副理事长，北京应用统计学会会长，国际数理统计学会（中国分会）常务理事。曾获得教育部高等学校科学技术奖-自然科学奖二等奖；全国统计科学研究优秀成果奖一等奖等。