Yanwei Fu

School of Data Science, Fudan University.

yanweifu@fudan.edu.cn
ztwztq2006@gmail.com

  • Home
  • Dataset
  • Students
  • Papers

    • Yanwei Fu is a full professor (tenured) at the School of Data Science, Fudan University, Shanghai China.

    He is  Professor of Eastern Scholar, Shanghai (上海高校特聘教授 – 东方学者), 1000 young scholar (青千)，and ARC DECRA fellow.

    Research interests

    The core scientific problem I investigate centers on few-shot learning, which seeks to emulate the human ability of “learning to learn”—enabling models to acquire new concepts and generalize to unseen tasks from limited data. This line of research focuses on how to effectively integrate prior knowledge with scarce samples to achieve robust knowledge transfer and generalization.

    While large-scale models such as Transformers and Diffusion networks have achieved remarkable success in 2D perception, they still encounter significant challenges under data-sparse conditions, particularly in 3D reconstruction and multi-modal learning. Addressing these limitations drives my exploration of broader yet closely related applications, including vision-guided robotic learning and fMRI-based neural decoding, where data scarcity and generalization remain fundamental bottlenecks. I’m particularly interested in these topics recently,

    1, Few-Shot Learning Theory and Algorithms: Large models often rely on massive labeled datasets, yet data collection is costly and prone to bias and noise. My research develops meta-learning and statistical frameworks—drawing on extreme value theory, causality, and sparsity—to enable models to learn and generalize effectively from limited samples.

    2, Few-Shot 3D Reconstruction: 3D reconstruction typically requires dense viewpoints and extensive data, while real-world acquisition remains limited. I integrate prior knowledge, neural networks, and optimization methods to recover accurate 3D structures from sparse observations, advancing practical applications in this domain.

    3, Multi-Modal Few-Shot Learning: Learning from multi-modal data—combining images, videos, audio, and sensor signals—requires efficient cross-modal fusion under data scarcity. I design frameworks that enhance generalization across modalities, enabling robots, for example, to learn 6DoF pose estimation and grasping from only a few demonstrations.

    Timeline

    Please donot contact me via my QMUL email address, which has been suspended.

    关于本组招生：大数据学院科学硕士及直博生是统一招生（而不是分配给导师名额）。所以感兴趣我们组的话，可以直接去申请拿到学院的offer，再联系我即可。申请普博或者Oversea students想来我们组做summer intern的话，可以提前联系我。 大家有兴趣联系我们，可以email联系我。 有兴趣做statistical sparsity的本科同学，可以直接联系孙鑫伟老师，我们可以一起合作。

    **注意：本组2025年入学的硕士已经招满，没有名额了。 **

    Recent News

    Media Interview