Diyuan Wu
Welcome to my homepage! My name is Diyuan Wu (吴迪远 in Chinese), I'm currently a PhD student in Institute of Science and Technology Austria (ISTA) starting from fall, 2022, fortunately working with Professor Marco Mondelli. Prior to my PhD study, I received a master degree in communication system from EPFL, Lausanne, Switzerland in 2022, and bachelor degree in information and communication engineering from UESTC(电子科大), Chengdu, China in 2019. Please find my lastest CV here.
I mainly work on machine learning theory. Specificly, my research works aim to: (1) theoretically understand experimental phenomenons and (2) use theoretical insights to improve practical utility and efficiency. I'm particularly interested in problems with high-dimensional data/models. Currently, I'm mainly working on:
- Feature learning: how neural networks map data to features and learn features from data.
Lastest updates
2025-05 New preprint on arXiv "Attention with Trained Embeddings Provably Selects Important Tokens." (Joint with Alex, Samet and Marco).
2025-05 Our paper "Neural Collapse Beyond the Unconstrained Features Model: Landscape, Dynamics, and Generalization in the Mean-Field Regime." (joint with Marco) accepted as spotlight paper at ICML 25.
Contact
I'm always open to discussions and collaborations on any related research topics. Please be more than welcomed to reach out by:
Email: diyuan.wu@ist.ac.at