Yuxin Wu

Wu works on foundation models at Moonshot AI after earlier work at Google Brain and Facebook AI Research. He earned a bachelor’s degree in computer science at Tsinghua in 2015 and a master’s in computer vision at Carnegie Mellon in 2016. His projects combine research on learning methods with software that makes experiments and deployment easier to build.[1]

Normalize without depending on a training batch

Group Normalization divides channels into groups and computes normalization statistics within each group. Unlike batch normalization, its calculation does not depend on other examples in a minibatch. Wu and Kaiming He study its behavior with different batch sizes and apply it to classification, detection, segmentation and video recognition. This addresses settings where large images or other memory demands limit the number of examples trained together.[2]

Learn visual representations through a changing dictionary

Wu coauthored Momentum Contrast, which frames contrastive representation learning as matching a query against encoded keys. A queue maintains a large dictionary of keys, while a momentum-updated encoder keeps their representations sufficiently consistent as training changes the model. The paper evaluates the resulting representations on image classification and transfers them to detection and segmentation tasks.[3]

Provide reusable computer-vision infrastructure

Wu created Detectron2, a PyTorch-based library for detection and related vision tasks. Its public code and model collection support research and production use across detection and segmentation architectures. He also developed Tensorpack, a TensorFlow training interface with data pipelines, trainers and paper reproductions, reflecting the same interest in research infrastructure.[1][4][5]