Yuxin Wu

Career

Wu worked on computer vision at Facebook AI Research and on foundation models at Google Brain. Wu now builds foundation models at Moonshot AI, the company behind Kimi.¹

Research

Wu's research at FAIR covered normalization, representation learning and segmentation. With Kaiming He, Wu wrote "Group Normalization" (ECCV 2018), and Wu co-authored Momentum Contrast (MoCo, CVPR 2020) for unsupervised visual representation learning, PointRend (CVPR 2020) for image segmentation and "Feature Denoising for Improving Adversarial Robustness" (CVPR 2019).¹⁴³⁵

Earlier work includes DoReFa-Net (2016) on training low-bitwidth networks and reinforcement learning agents for first-person shooter games. At Moonshot AI, Wu is listed with the Kimi Team on the 2025 Kimi K2 report, which describes an open-weight agentic language model.¹⁶⁷

Open-source work

Wu created Detectron2, FAIR's library for detection and segmentation, which supports training and deployment for Meta products and received the Mark Everingham Prize at ICCV 2021. Wu also wrote Tensorpack, a neural network training interface for TensorFlow focused on speed and flexibility.¹²

Work and contributions

• Detectron2 Open-source code

  Wu's site says Wu created Detectron2; the repository lists Wu first in its citation entry and releases the library under the Apache 2.0 license.¹²

• MoCo (Momentum Contrast) Published research

  Co-author of "Momentum Contrast for Unsupervised Visual Representation Learning" (2019) with Kaiming He, Haoqi Fan, Saining Xie, and Ross Girshick.³

• Group Normalization Published research

  First author, with Kaiming He, of "Group Normalization" (ECCV 2018), which normalizes features within groups of channels so that accuracy holds up with small batch sizes; the paper received an ECCV 2018 best paper honorable mention.⁴¹

Timeline

1. Jun 20, 2016Co-authored DoReFa-Net, on training low-bitwidth convolutional neural networks.⁶
2. Mar 22, 2018Released "Group Normalization" as first author with Kaiming He.⁴
3. Nov 13, 2019Co-authored "Momentum Contrast for Unsupervised Visual Representation Learning" (MoCo).³
4. Dec 17, 2019Co-authored "PointRend - Image Segmentation as Rendering".⁵
5. 2021Detectron2, created by Wu, received the Mark Everingham Prize at ICCV 2021.¹
6. Jul 28, 2025Listed with the Kimi Team on the Kimi K2 open agentic intelligence report.⁷

Other documented roles

Roles the sources document that are not confirmed as current.

• Worked on foundation models at Google Brain, Google Brain¹
• Worked on computer vision at Facebook AI Research, Meta (Facebook AI Research)¹

Honors and recognition

• ECCV Best Paper Honorable Mention (Group Normalization) · European Conference on Computer Vision, 2018¹
• CVPR Best Paper Nomination (Momentum Contrast) · Conference on Computer Vision and Pattern Recognition, 2020¹
• PAMI Mark Everingham Prize (Detectron2) · IEEE PAMI Technical Committee, 2021¹

Where they work

• Moonshot AI 2 people in the wiki

Related people

• Xu TanAlso at Moonshot AI
• Dong YuML frameworks and libraries · Language models
• Michael AuliLanguage models · ML frameworks and libraries
• Luis LastrasLanguage models · ML frameworks and libraries
• Fabian PedregosaML frameworks and libraries · Language models
• Jyoti AnejaComputer vision · Language models
• Luke Melas-KyriaziLanguage models · Computer vision
• Clément FarabetLanguage models · ML frameworks and libraries
• Gabriel SynnaeveLanguage models · ML frameworks and libraries
• Tong HeML frameworks and libraries · Computer vision