He Kaiming officially announced that he has joined Google DeepMind.

According to a report by Zhidx on June 26th, recently, He Kaiming, a top figure in computer vision and a professor at MIT, has joined Google DeepMind. He updated his latest position on his personal homepage - Distinguished Scientist (Part - time) at Google DeepMind. Meanwhile, He Kaiming still retains his tenured faculty position at the Department of Electrical Engineering and Computer Science (EECS) at MIT.

He Kaiming is one of the proposers of the Deep Residual Network (ResNet). The ResNet paper "Deep Residual Learning for Image Recognition" he published as the first author is the most - cited paper in the 21st century.

The residual connections proposed in the paper are ubiquitous in modern deep - learning models, including Transformers, AlphaGo Zero, AlphaFold, and almost all current generative AI models. As of May this year, the total citation count of He Kaiming's various publications has exceeded 700,000.

Before joining MIT in 2024, He Kaiming was always active in both the industry and academia. He successively served as a research scientist at Microsoft Research Asia (MSRA) and Facebook AI Research (FAIR). He holds a bachelor's degree from the Department of Physics, Tsinghua University, and a doctorate in Information Engineering from The Chinese University of Hong Kong. During his undergraduate years, he interned at the Visual Computing Group of Microsoft Research Asia, under the guidance of Sun Jian, a top figure in computer vision. At The Chinese University of Hong Kong, he studied under Tang Xiaoou, the founder of the Multimedia Laboratory at The Chinese University of Hong Kong and the founder of SenseTime.

In addition to ResNet, He Kaiming has also published many research results of great academic value that have had a profound impact on AI and computer vision.

In 2009, while at The Chinese University of Hong Kong, He Kaiming proposed the "image dehazing algorithm" in his first published academic paper. As soon as this paper was published, it won the Best Paper Award at the top computer vision conference CVPR that year. He Kaiming also became the first Chinese to receive this honor in CVPR history.

In 2015, the ResNet proposed by He Kaiming when working at Microsoft Research Asia won the championship in the ImageNet image recognition competition, and the related paper won the Best Paper Award at CVPR in 2016.

During his time at Facebook AI Research, He Kaiming also made important contributions in the field of image segmentation. He published two important papers, Mask R - CNN and Faster R - CNN, as the first author and the second author respectively. The related research has improved the accuracy and efficiency of image segmentation to a new level, and the Mask R - CNN paper won the Best Paper Award at ICCV in 2017.

He Kaiming has also won many famous awards, such as the PAMI Young Researcher Award in 2018, the Best Paper Honorable Mention at ECCV 2018 and CVPR 2021, and the Everingham Award at ICCV 2021.

Currently, neither Google DeepMind nor He Kaiming himself has disclosed the specific arrangements after his joining. However, we can learn about the research directions he considers academically valuable from his recent sharing at the CVPR and NeurIPS conferences.

Since AlexNet, recognition models have generally achieved end - to - end training and inference. However, the current mainstream generative models are conceptually similar to "hierarchical training" and usually involve multi - step inference and calculation. The team led by He Kaiming published the theoretical framework of the single - step generative model MeanFlow in 2025. In the future, he may continue to explore frameworks suitable for end - to - end generative modeling.

At the same time, He Kaiming also said that recognition and generation are two sides of the same coin. Recognition is the "flow" from data to embedding, while generation is the "flow" from embedding to data. In the future, an integrated framework for recognition and generation may also become one of his important research directions.

When sharing at the NeurIPS conference in 2024, He Kaiming emphasized: "The future is the real test set". He advocates that researchers should focus on new, unseen data, new configurations, new use cases, and new scenarios to reduce "overfitting" in research.

At the same conference, He Kaiming said that the essence of research is to find "surprises". After joining Google DeepMind, we may expect him to bring more surprising scientific research results.

This article is from the WeChat official account "Zhidx" (ID: zhidxcom). Author: Chen Junda, Editor: Li Shuiqing. It is published by 36Kr with authorization.