Canran Xiao | 肖粲然

About Me

Hello! I am Canran Xiao, a postdoctoral fellow at Sun Yat-sen University. My postdoctoral mentor is Prof. Li Shen. I received my Ph.D. from Central South University, advised by Prof. Xiaohong Chen and Prof. Yongmei Liu.

My research is centered on data intelligence (i.e., data-centric machine learning, where data is the key driver of intelligence). I study how to make complex data easier for foundation models to learn from, evaluate, and use reliably. My work focuses on dynamic data learning, heterogeneous data learning, and data attribution and valuation. I have published over twenty papers in CCF-A venues, including NeurIPS, ICLR, ICML, AAAI, WWW, CVPR, and ACM MM.

I welcome collaborations on data-centric learning, especially on solving data problems in foundation models.

News

Selected Publications

Selected papers from my Google Scholar. * denotes corresponding author.

1. ICML 2026

   Influence-Disentangled Federated Training: Learning Models That Are Easy to Unlearn CCF-A Poster

   C Xiao, Q Chen, L Hou

   Forty-third International Conference on Machine Learning (ICML 2026)

   Highlight: separates data influence in federated training so learned models can remove data contributions more cleanly.

2. ICML 2026

   In-Context Learning as Rate–Distortion Optimization CCF-A Poster

   J Zhang, C Li, C Xiao*

   Forty-third International Conference on Machine Learning (ICML 2026)

   Highlight: studies how context data can be compressed and selected to improve foundation-model inference.

3. ECCV 2026

   Interference-Aware Continual Vision-Language Learning via Instance-Level Expert Routing CCF-B

   C Zhang, T Xu, F Shen, C Xiao*

   The 19th European Conference on Computer Vision (ECCV 2026)

   Highlight: routes evolving vision-language data to instance-level experts to reduce interference during continual learning.

4. WWW 2026

   Prototype-Aligned Federated Soft-Prompts for Continual Web Personalization CCF-A Oral

   C Xiao, L Hou

   The ACM Web Conference 2026 (WWW 2026)

   Highlight: handles continuously arriving user data while keeping personalized web models adaptive across clients.

5. ICLR 2026

   Reversible Primitive-Composition Alignment for Continual Vision-Language Learning CCF-A Poster

   C Xiao, T Xu, S Ma, Y Jiang, H Gao, Y Wu

   The Fourteenth International Conference on Learning Representations (ICLR 2026)

   Highlight: supports data that arrives over time by aligning primitive and compositional knowledge in vision-language streams.

6. ICLR 2026

   Path Matters: Unveiling Geometric Implicit Bias via Curvature-Aware Sparse View Optimization CCF-A Poster

   C Xiao, L Fan, Y Li, J Tang, P Yu

   The Fourteenth International Conference on Learning Representations (ICLR 2026)

   Highlight: improves sparse-view data utility by choosing camera paths that better cover high-curvature geometric information.

7. CVPR 2026

   Affordance-First Decomposition for Continual Learning in Video-Language Understanding CCF-A Poster

   M Xu, H Liu, N Peng, Q Chen, C Xiao*

   The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2026)

   Highlight: decomposes incoming video-language data so models can absorb new visual evidence without forgetting old data.

8. AAAI 2026

   From Points to Coalitions: Hierarchical Contrastive Shapley Values for Prioritizing Data Samples CCF-A Oral

   C Xiao, J Dou, Z Lin, Z Ke, L Hou

   The Fortieth Annual AAAI Conference on Artificial Intelligence (AAAI 2026)

   Highlight: estimates which data samples contribute the most value, helping prioritize data before training.

9. NeurIPS 2024

   Confusion-resistant federated learning via diffusion-based data harmonization on non-iid data CCF-A Poster

   X Chen, C Xiao*, Y Liu

   The Thirty-eighth Conference on Neural Information Processing Systems (NeurIPS 2024)

   Highlight: harmonizes heterogeneous client data distributions so federated learning can work better on non-IID data.

10. NeurIPS 2024

    Swift sampler: Efficient learning of sampler by 10 parameters CCF-A Poster

    J Yao, C Li, C Xiao*

    The Thirty-eighth Conference on Neural Information Processing Systems (NeurIPS 2024)

    Highlight: learns how to sample informative data efficiently, reducing the cost of deciding which data to use.

Academic Service