Speaker: Xiaotian (Max) Han

Date: Feb 16, 11:45am–12:45pm

Abstract: In this talk, I will present my research on resource-efficient machine learning techniques for graph neural networks (GNNs) and beyond. These techniques aim to reduce the computational resources required by these models, making them more practical for real-world applications. i) I will discuss accelerating the training and inference of GNNs by connecting them with multilayer perceptrons (MLPs). I discovered that the weights between the GNN and MLP layers are transferable. Thus I propose to transfer the weights of MLP to initialize the GNN. This significantly reduces the training time of the GNN. Additionally, I creatively reformulated GNNs into a form of Mixup that can be implemented through an MLP, further improving efficiency. ii) I will discuss reducing the cost of data collection for graph learning. Since the graph data are expensive due to the complex data structure, I designed a graph data augmentation method to augment graph data for the graph classification task. Due to the different number of nodes across graphs, instead of directly mixing up graphs, we proposed mixing graph generators (i.e., graphons) instead. By doing this, we augment graph data and reduce the cost of graph learning. My research aims to address computational and efficiency challenges in large models, thereby enabling their use in more real-world applications. I hope to democratize state-of-the-art AI capabilities for high-impact societal use cases even with limited resources.

Biographical Sketch: Xiaotian (Max) Han is a Ph.D. candidate in computer science at Texas A&M University, advised by Dr. Xia (Ben) Hu. His research interests lie in artificial intelligence, machine learning, and data science, with a focus on designing resource-efficient deep learning methods for computationally restricted environments. He aims to democratize cutting-edge machine learning for high-impact societal applications with limited resources. He has published over 20 papers, including 11 first-authored papers, in top-tier machine learning conferences and journals such as ICML, ICLR, NerIPS, TMLR, WWW, KDD, IJCAI, AAAI, TKDE, etc. He has also served as reviewer for these top conferences. He was the recipient of the Outstanding Paper Award at ICML 2022 as the first author.

Location and Zoom link: 307 Love, or https://fsu.zoom.us/j/94676311356?pwd=MUtkQmhRRlU4YVVNRW90OGpYZWZuQT09