Most famous AI scientists pick a lane — the university podium or the corporate lab. Shuicheng Yan (颜水成) kept crossing the line so often that the line stopped meaning anything to him. In 2023 he added one more border crossing, leaving Singapore's academic establishment to help run a Chinese public company's push into generative AI.
Key takeaways
- Shuicheng Yan (颜水成) is a computer-vision and machine-learning scholar, an Academician of the Singapore Academy of Engineering and a Fellow of AAAI, ACM, IEEE, and IAPR.
- His team's "Network In Network" paper introduced the 1×1 convolution and global average pooling that became standard building blocks in GoogLeNet and ResNet.
- He moved between industry and academia for two decades — 360 Group, YITU, and Sea Group's Sea AI Lab — before joining Kunlun Tech (昆仑万维) in September 2023 as co-CEO of SkyWork AI (天工智能) alongside founder Zhou Yahui.
- His recurring thesis, stated when he joined Kunlun, is that research, development, and product are "one complete chain" in which none of the three can be missing.
A mathematician who rewired convolutions
Yan's grounding is mathematical. He earned a bachelor's degree in 1999 and a doctorate in 2004 from the School of Mathematical Sciences at Peking University, then did postdoctoral work at the Chinese University of Hong Kong's multimedia lab — under Tang Xiaoou, another foundational figure in Chinese computer vision — and at the University of Illinois at Urbana-Champaign.
In 2007 he joined the National University of Singapore and founded its Machine Learning and Computer Vision Lab. His most cited early contribution is "Network In Network" (NIN), the work that popularized the 1×1 convolution and global average pooling. Those ideas look mundane now, but they were later absorbed into GoogLeNet and ResNet — the architectures that carried deep learning through the 2010s. For a researcher, that is a rare kind of influence: not a product, but a brick everyone else builds with.
The academic-industrial commute
What distinguishes Yan is not a single breakthrough but the frequency with which he has stepped into operating roles:
- 360 Group (2015–2019) — vice president, chief scientist, and head of the AI research institute, his first major plunge into industry.
- YITU Technology (2019–2020) — chief technology officer, working on commercializing fundamental vision algorithms.
- Sea Group (2020–2023) — group chief scientist and founder of Sea AI Lab, the research arm behind the Southeast Asian internet group.
- Kunlun Tech (from September 2023) — co-CEO of SkyWork AI (天工智能) with founder Zhou Yahui, and director of the Kunlun 2050 Global Research Institute.
That résumé reads like a map of where Chinese-language AI commercialization happened — security and consumer internet, then founders'-led model labs, then public-company scale.
The Kunlun move
Kunlun Tech (昆仑万维) is a Shenzhen-listed internet company that, in 2023, declared an "All in AGI" strategy and reorganized its AI work under SkyWork AI, incorporated in Singapore. Yan's September 2023 appointment as co-CEO of SkyWork AI, announced in the company's own English release, was meant to signal that the push would be research-led, not just product-led.
When he joined, Yan explained the logic in remarks carried by Shanghai Securities News: in general artificial intelligence, he said, research, development, and product form a complete chain, with none of the three dispensable, and only by fully connecting all three can research create its greatest value. He added that foundation models (大模型) were "the current wave" — the most exciting moment in AI. It is a tidy summary of why a tenured-style academic would keep returning to the factory floor.
What the "complete chain" means in practice
Yan's bet is that the gap between a good paper and a usable product is where most AI value is lost. SkyWork AI under his co-leadership has spanned text, music, video, and agent tooling — the company's Tiangong (天工) model family and its Skywork Super Agents workspace are the consumer-facing ends of that chain. The research-institute arm, the 2050 Global Research Institute, was set up to chase longer-horizon questions such as next-generation foundation models and agent evolution.
The risk in this model is familiar: a research star lends prestige, but translating that into shipped, defensible products is a different discipline. Yan's entire career has been an argument that the two can be run as one loop rather than as sequential handoffs.
Honest limitations
This profile draws on Kunlun Tech's official English announcement of Yan's appointment, Shanghai Securities News coverage of his joining remarks, and standard biographical listings of his academic honors and prior roles. Counts of his publications vary by source (commonly cited at more than 600 papers with an H-index in the high double digits); we describe his output qualitatively rather than asserting a single precise figure. The influence of "Network In Network" on later architectures such as GoogLeNet and ResNet is well established in the deep-learning literature, but the exact lineage of specific layers is a matter of citation practice rather than a formal dependency we can prove line by line. SkyWork AI's product traction and model benchmarks are reported by Kunlun Tech itself and were not independently audited for this piece. Yan did not grant a new interview for this article, so his views are represented through previously published, attributable statements.
What readers can do now
- Judge "research-led" AI companies by their pipeline, not their press releases: check whether papers from their labs actually ship into the product you use.
- When evaluating a foundation-model startup, ask whether it has a named research leader with a real publication record or only a marketing one — Yan's career is a useful contrast case.
- If you experiment with SkyWork AI's Tiangong (天工) tools or Skywork agents, note where the output is genuinely novel versus repackaged — that boundary reveals how mature the "complete chain" really is.