On a winter afternoon in 2015, a young researcher in a Beijing office hit a wall that most people never hear about: make a neural network deeper, and it simply stopped learning. He wrote a short paper with three colleagues. A decade later, that paper sits underneath nearly every image model, translation engine, and chatbot in daily use.
Kaiming He (何恺明) is not a household name, but his work is. The architecture he co-authored, ResNet, is the reason deep learning could scale past a couple of dozen layers — and why the large model (大模型) era was even possible.
The problem nobody could name
Before 2015, adding layers to a convolutional network usually made it worse. Researchers called it "degradation": the model would train, but accuracy plateaued and then fell, even on the training data itself. It was not overfitting. The network was effectively unable to learn the identity mapping it needed.
He, then a researcher at Microsoft Research Asia (MSRA) in Beijing, reframed the task. Instead of asking each stack of layers to learn a target directly, he asked it to learn only the difference from its input — a residual. A shortcut connection carried the original input forward and added it back. The layers now only needed to nudge a signal, not rebuild it from scratch.
That single idea let teams train networks 50, 101, and 152 layers deep. An ensemble of ResNets won the 2015 ImageNet challenge across every track, posting a 3.57% top-five error — below the estimated human rate.
What ResNet actually changed
The paper, "Deep Residual Learning for Image Recognition," was posted as a preprint in December 2015 and presented at CVPR in 2016. Its influence spread far past vision:
- The residual connection (残差连接) became a standard building block in Transformers, the architecture behind today's language models.
- It underpins AlphaGo Zero and AlphaFold, where deeper, stable networks were essential.
- According to an analysis published by Nature in 2025, it is the most-cited research article of the 21st century, with citation counts reaching into the hundreds of thousands.
- He's own publication record, per his MIT profile, passed 700,000 citations by May 2025.
ResNet did not win because it was clever on a benchmark. It won because it removed a ceiling. Once training very deep networks was routine, everyone else's models got deeper too.
A career across three labs
He's path is unusual for its consistency. Born in Guangzhou in 1984, he topped the Guangdong college entrance exam before studying at Tsinghua University, then earned his PhD at the Chinese University of Hong Kong under Xiaoou Tang (汤晓鸥). His very first paper — on removing haze from a single image — won a CVPR best-paper award in 2009, the first time a Chinese or Asian researcher took that honor.
- 2011–2016: researcher at MSRA, where ResNet was born.
- 2016–2024: research scientist at Facebook AI Research (FAIR), where he produced Mask R-CNN, MoCo, and MAE.
- February 2024: joined MIT as an associate professor in electrical engineering and computer science.
- July 2024: named the Douglas Ross Career Development Professor.
- June 2025: granted tenure at MIT and, in the same month, took a part-time role as a distinguished scientist at Google DeepMind.
In 2023 he shared the Future Science Prize (未来科学大奖) in mathematics and computer science — a US$333,000 award — with Xiangyu Zhang, Shaoqing Ren, and the late Jian Sun, for introducing deep residual learning.
Why this matters for China's AI
He now works in the United States, but the work that defined his reputation was done in Beijing, inside a Chinese research lab. That detail matters for an observer of China's AI talent: the country produces researchers whose ideas become global infrastructure, even when the researchers themselves migrate.
Two readings follow. The optimistic one is that China's training pipeline — Tsinghua, CUHK, MSRA — can surface world-shaping scientists. The cautious one is that the rewards for that talent increasingly sit abroad. He's "academia plus industry" arrangement at MIT and DeepMind is exactly the kind of setup Chinese labs have been trying to replicate.
Honest limitations
This profile relies on publicly reported biographical and citation data, including He's MIT profile, encyclopedia entries, and media coverage. Precise citation counts shift constantly and vary by database; the Nature "most-cited of the century" claim refers to a specific 2025 analysis and should not be read as a permanent ranking. I have not interviewed He, and some details of his early career come from Chinese-language profiles that I could not independently cross-check against primary documents. The link between ResNet and any specific commercial product is causal at the architecture level, not a claim that He built those products.
What readers can do now
- If you build or buy AI systems, check whether your vision or multimodal stack uses a ResNet-family backbone; understanding that lineage explains a lot about why models behave as they do.
- For students, He's paper is a model of restraint: a small conceptual change, clearly explained, that unlocked a field. Read the 2015 preprint before the hot takes.
- For observers of China's talent flow, track where flagship researchers land and what domestic labs offer in return — that competition is the real story behind the citation counts.