Optimal Transport-driven Design of Graph Neural Networks

报告题目：Optimal Transport-driven Design of Graph Neural Networks

报告人：许洪腾   副教授（中国人民大学 高瓴人工智能学院）

主持人：王祥丰   副教授

报告时间：2024年4月1日（星期一）10:00-11:00

报告地点：华东师范大学普陀校区理科楼B504

报告摘要：

Nowadays, artificial intelligence techniques, like deep learning, have achieved encouraging performance on high-dimensional data representation and generation tasks. However, these successes depend on highly specialized neural networks, whose designs are empirical and often inapplicable to scenarios requiring high interpretability, such as medicine and transportation. Focusing on the above challenges, we would like to build an optimal transport-driven paradigm for neural network design, focusing state-of-the-art graph neural networks and improving their key modules. In particular, we show that the feed-forward computation of typical graph neural network layers, including global pooling and message-passing, can be reformulated as solving optimal transport (OT) problems, resulting in new model architectures and learning paradigms. Experiments demonstrate that the proposed OT-driven design leads to more generalizable and interpretable graph neural networks and boosts their performance consistently in various learning tasks.

报告人简介：

       许洪腾，中国人民大学长聘副教授、博士生导师。2017年博士毕业于美国佐治亚理工学院，2017年至2018年担任杜克大学博士后研究员，2018-2020年担任美国InfiniaML公司高级研究员，2021年就职于中国人民大学高瓴人工智能学院。研究方向为面向结构化数据的机器学习及其应用，主要研究基于计算最优传输理论和算法的复杂数据分析、建模、预测、生成及控制技术，及其在生物医药领域的应用。在ICML、NeurIPS、AAAI、IJCAI、TPAMI等人工智能国际顶级会议和期刊上发表论文40余篇。2021年入选国家级高层次青年人才项目。