Welcome to Zitao Shuai’s Home Pages
Zitao Shuai 帅子滔
Biography
Currently, I’m a master’s student in the EECS department at the University of Michigan. Currently, I’m advised by Prof.Liyue Shen in the University of Michigan and I’m also collaborate with Prof.Kun Kuang’s group in Zhejiang University. I got my bachelor’s degree in Computer Science and my bachelor’s degree in Finance from Zhejiang University. And I’m also a member of CKC Honors College. During my undergraduate, I was advised by Prof.Kun Kuang in the Department of Computer Science and Technology of ZJU.
My current research interests are Multi-modal Learning
/Trustworthy AI
/AI for Science
currently.
Publication
[1] Zhang F*, Shuai Z*, Kuang K, et al. Unified fair federated learning for digital healthcare[J]. Patterns, 2024, 5(1). [Subjournal of Cell Press]
[2] Zitao Shuai and Liyue Shen. 2024. Mitigating Heterogeneity in Federated Multimodal Learning with Biomedical Vision-Language Pre-training. (2024). arXiv:2404.03854 [cs.LG]
Research
In-Context Medical Segmentation
- Timeline: Fed 2024-March 2024
- Institution: Biomedical-AI Lab, UMich
- Instructor: prof.Liyue Shen
- Motivation: Efficient In-Context Medical Segmentation with Meta-driven Visual Prompt Selection
- Product: A MICCAI 2024 submitted paper, 3th author.
Heteroneneity on Biomedical Multi-Modal Pre-training
- Timeline: Nov 2023-Jan 2024
- Institution: Biomedical-AI Lab, UMich
- Instructor: prof.Liyue Shen
- Motivation: Mitigating Heterogeneity in Federated Multimodal Learning with Biomedical Vision-Language Pre-training
- Product: A ICML 2024 submitted paper, 1st author.
Longitudinal Study of Biomedical Vision-Language Pre-training
- Timeline: Oct 2023-Nov 2023
- Institution: Biomedical-AI Lab, UMich
- Instructor: prof.Liyue Shen
- Motivation: Biomedical Vision-Language Pretraining typically pairs single images with text, but overlooks temporal and localized changes in medical data, thus limiting its effectiveness for complex real-world clinical applications.
- Product: A CVPR 2024 submitted paper, 2nd author.
Unified Fair Federated Learning for Digital Healthcare
- Timeline: May 2023-Aug 2023
- Institution: Institute of Artificial Intelligence, ZJU
- Instructor: prof. Kun Kuang
- Motivation: To ensure fairness in federated learning with healthcare institutions, we employ distributionally robust optimization to achieve consistent and high-performing federated models among diverse demographics.
- Contribution: presents a unified framework encompassing various fairness metrics in federated learning and devises an efficient and collaborative machine learning procedure for its implementation.
- Product: A paper published in Cell Press Pattern 2023, co-first author.
Continual Domain Generalization
- Timeline: Aug 2022-Jan 2023
- Institution: Institute of Artificial Intelligence, ZJU
- Instructor: prof. Kun Kuang
- Motivation: Tackle the challenging problem of deploying the Domain Generalization model in the continual data stream.
- Method: Learn domain invariance from the data stream by fusing the new and old knowledge flexibly in a contrastive support space. Specifically, train a prior-optimized variational auto-encoder using a reconstruction loss to initialize a semantic preservation latent space; adopt contrastive learning to construct a contrastive support space, which is used to obtain exemplar data as memory to overcome the forgetting problem during stream training.
- Contribution: Propose a new task and provide a novel method to solve this problem.
- Product: A NIPS 2023 submitted paper, 1st author.
- Timeline: May-Sept 2022
- Institution: CSC Lab, UC Davis
- Instructor: prof. Hao-Chuan Wang
- Motivation: Investigate if peers’ thinking comments can help nudge individuals on higher-order thinking processes.
- Method: Complete user experiments on the well-designed prototypes based on our assumptions.
- Role: Implement the Deep Thinking Map prototype based on the Concept Map developed in our lab; help organize pilot experiments; literature surveying
Spatio-temporal Data Compression
- Timeline: Jul-Sept 2021
- Institution: Center of Data Science, ZJU
- Motivation: Solve the spatiotemporal data compression problem with the techniques for graph evolution.
- Method: Combine a particle-swarm-optimization-based graph clustering algorithm and a compression algorithm (a well-designed data structure ) for dynamic-evolution spatiotemporal data.
- Role: reproduce two baseline papers with C++ codes and run several large-scale datasets to conduct the case study.
Awards
- Outstanding Graduate of Zhejiang University
- Meritorious Winner in Mathematical Contest in Modeling
Notes & ideas
notes for paper
notes for presentation