Kaiming He

Overview

Kaiming He is a computer-vision and deep-learning researcher. He is the first author of "Deep Residual Learning for Image Recognition", the 2015 paper that introduced ResNet, written with Xiangyu Zhang, Shaoqing Ren and Jian Sun at Microsoft Research.^[3]

Education and early work

He placed first in total score in the 2003 Guangdong provincial college entrance exam, the gaokao.^[2] He earned a B.S. from Tsinghua University in 2007 and a PhD from the Chinese University of Hong Kong in 2011, where his advisor was Xiaoou Tang and his thesis was on single-image haze removal using the dark channel prior.^[2]

That thesis topic produced his first widely noticed result. The dark channel prior is a statistical observation about ordinary outdoor photographs: in most local patches at least one colour channel is very dark, and haze lifts exactly those dark values. Comparing a hazy image against that expectation gives an estimate of how much haze covers each point, which in turn allows the haze to be subtracted from a single picture, without a second photo or extra information about the scene. The work won a CVPR Best Paper Award in 2009.^[2]

ResNet

The arXiv preprint of "Deep Residual Learning for Image Recognition" was submitted on 10 December 2015, listing Kaiming He, Xiangyu Zhang, Shaoqing Ren and Jian Sun as authors.^[3]

Its idea is easy to state. Instead of asking a stack of layers to learn a target mapping directly, each block learns a residual, the difference between the target and the block's input, and the input is added back through a shortcut connection. A block can then do nothing at all by driving the residual to zero, so adding depth no longer makes training worse, and very deep networks become trainable.

Residual connections turned out to be far more general than image classification. His homepage describes the ResNet paper as the most-cited paper of the twenty-first century according to a Nature analysis, and notes that residual connections are used in Transformers, AlphaGo Zero, AlphaFold and virtually all generative-AI models.^[1]

Detection, segmentation and self-supervised learning

He also worked on object detection and segmentation, the tasks of locating objects in an image and outlining them pixel by pixel, with Faster R-CNN (NeurIPS 2015) and Mask R-CNN (ICCV 2017). In self-supervised learning, where a model learns representations from unlabelled data by solving a task constructed from the data itself, he worked on MoCo and MAE.^[1]

Career

He was a researcher at Microsoft Research Asia from 2011 to 2016, a research scientist at Facebook AI Research (FAIR) from 2016 to 2024, and from 2024 an associate professor with tenure in MIT's Department of Electrical Engineering and Computer Science. He also works part-time as a Distinguished Scientist at Google DeepMind.^[1]

Recognition

His awards include CVPR Best Paper Awards in 2009, for dark channel prior haze removal, and in 2016, for ResNet; the ICCV 2017 Best Paper Award, the Marr Prize, for Mask R-CNN; the PAMI Young Researcher Award in 2018; test-of-time awards at NeurIPS 2025 and CVPR 2026; and the 2023 Future Science Prize, shared with three collaborators, for introducing deep residual learning.^[2] His homepage reports more than 700,000 citations as of May 2025.^[1]