报告题目: Towards better robustness and generalisation of metric learning methods

报告人：杨笑尘（英国格拉斯哥大学）

报告时间：2023年4月14日17:45-18:30

报告地点：文波楼201

摘要: Many statistical and machine learning methods, such as K-nearest neighbo urs and K-means, heavily depend on the distance measure between data points. As each task has its own notion of distance, metric learning has been proposed, w hich formulates an optimisation problem to learn a distance metric such that sem antically similar samples are pulled together while dissimilar samples are pushed away. In this presentation, I will present two algorithms to learn the distance met ric, which can be applied to continuous data and categorical data respectively. M ore specifically, the first one is motivated to improve robustness of metric learnin g algorithms against adversarial examples, and the second one takes advantage o f adversarial training, a widely used strategy to improve adversarial robustness, t o mitigate feature ambiguity. In addition, I will give a brief overview of recent dev elopment in deep metric learning and an application of metric learning in renewa ble systems.

报告人简介：杨笑尘，英国格拉斯哥大学数学与统计学院的讲师。于2020年获得伦敦大学学院统计科学博士学位，主要研究方向：静态机器学习和图像分析。目前，致力于距离度量研究，是《神经计算》编辑委员。