报告人：Prof. Pan Li
时 间：2022/11/3 09:00 - 10:00
Graph-structured data and point cloud data are ubiquitous across scientific fields. Geometric deep learning (GDL) has recently been widely applied to solve prediction tasks with such data. GDL models are often designed with complex equivariant structures to preserve geometric principles and are thus hardly interpretable. Moreover, GDL models also risk capturing spurious correlations between the input features and labels, which poses many concerns to scientists who aim to deploy these models in scientific analysis and experiments. In this talk, I will introduce our recent project on interpretable and trustworthy GDL models with applications with scientific data analysis. Our approach is based on a novel learnable randomness injection (LRI) mechanism, which is grounded by the information-bottleneck principle and can be applied to general GDL backbones. I will also introduce our recently established benchmarks with real-world applications in high energy physics and biochemistry to evaluate interpretable and trustworthy GDL models.
Pan Li joined the Purdue CS department as an assistant professor since 2020 Fall and is going to join Georgia Tech ECE department since 2023 Spring. He researches interest lies broadly in the area of machine learning and optimization on graphs. His recent works include algorithms and analysis of graph neural networks, hypergraph spectral theory and optimization, applications of graph machine learning in physics and design automation. Pan Li has got several awards including JPMorgan Faculty Award, Sony Faculty Innovation Award and Ross-Lynn Faculty Award.