My name is George Ma (Jiangyan Ma). I am an EECS PhD student at UC Berkeley, advised by Prof. Somayeh Sojoudi. Previously, I was an undergraduate student at Peking University, where I did research on graph learning at Prof. Yisen Wang’s lab. My email: georgem@stu.pku.edu.cn. My homepage: George Ma’s Homepage. I actively write blogs on Zhihu: George M’s Zhihu Homepage.
Undergraduate; Information and Computing Science
I studied at School of Electronic Engineering and Computer Science, Peking University. I majored in Applied Physics in the first two years of college and switched to Information and Computing Science thereafter.
PhD student; Electrical Engineering and Computer Science
Currently, I am an EECS PhD student at UC Berkeley.
Jiangyan Ma, Yifei Wang, Yisen Wang (2023). Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding. In Thirty-seventh Conference on Neural Information Processing Systems.
Jiangyan Ma, Emmanuel Bengio, Yoshua Bengio, Dinghuai Zhang (2023). Baking Symmetry into GFlowNets. In NeurIPS 2023 AI for Science: from Theory to Practice.
George Ma, Yifei Wang, Derek Lim, Stefanie Jegelka, Yisen Wang (2024). A Canonization Perspective on Invariant and Equivariant Learning. In Thirty-eighth Conference on Neural Information Processing Systems.
July 2022 – October 2022; Universal Graph Neural Networks without Message-Passing
In this research experience, I focused on studying the expressive power of MLP-based graph neural networks for graph-level tasks. I developed a novel MLP-based GNN called UBoN (Universal Bag of Nodes), specifically designed to achieve universal expressive power on graphs. I submitted my paper titled Universal Graph Neural Networks without Message Passing to ICLR 2023, although it was unfortunately rejected.
October 2022 – August 2023; Laplacian Canonization for GNN Spectral Embedding
During my research, I explored the concept of Laplacian eigenvectors as universal graph positional encodings. Although being universal, there are challenges associated with their universality, such as sign and basis ambiguity. To tackle these issues, I proposed an innovative pre-processing technique called Laplacian Canonization. I submitted my paper titled Laplacian Canonization: A Minimalist Approach to Sign and Basis Invariant Spectral Embedding to NeurIPS 2023, and it was accepted as a poster.
June 2023 – September 2023; Baking Symmetry into GFlowNets
I actively contributed to groundbreaking research as a member of Prof. Yoshua Bengio’s lab at Mila - Quebec AI Institute in Montréal. Leveraging my expertise in invariant networks, I was dedicated to incorporating symmetries into the graph construction process of GFlowNets. Preliminary experiments demonstrated promising performance enhancements. I submitted my paper titled Baking Symmetry into GFlowNets to the NeurIPS 2023 AI4Science Workshop, and it was accepted as an oral presentation.
June 2023 – May 2024; A Canonization Perspective on Invariant and Equivariant Learning
I collaborated with researchers from MIT to investigate the correlation between the canonizability of Laplacian eigenvectors and the expressive power of invariant networks (such as SignNet and BasisNet) that employ Laplacian eigenvectors as positional encodings. We established the theoretical framework of canonization, and applied it to solve the open problem of the expressive power of invariant networks with equivariance constraints. We also designed novel canonization algorithms for eigenvectors that are superior to prior work and even optimal. Our paper was accepted in NeurIPS 2024 as a poster.
July 2024 – Present; Local SAEs for Context Adaptation
I am collaborating with MIT researchers to explore the interpretability of large language models (LLMs). In our field, we use sparse auto-encoders (SAEs) to extract features from LLMs, which helps us understand and influence their behavior. However, training SAEs is costly, as we must start from scratch for each model, layer, and dataset. Since we often focus on specific contexts (like security or coding), we aim to design “local” SAEs that adapt to different contexts without needing re-training and can be applied to any model. We are preparing this work for submission to an upcoming machine learning conference.