Most people who use a chatbot never hear the name Shuicheng Yan (颜水成). But if you have ever wondered why a deep neural network can classify an image without a wall of fully-connected layers at the end, you have met his fingerprints. In 2013, Yan's team proposed Network In Network (NIN), which introduced global average pooling as a replacement for those layers — a small idea that quietly shaped how modern convolutional networks are built.
More than a decade later, Yan is still at the frontier, now focused on a different gap: giving voice agents a memory. In 2025 his team open-sourced VoiceMem, a "streaming dual-brain" memory framework for real-time speech agents. It is the latest turn in a career defined by constant movement between academia and industry.
From Peking University to a Singapore lab
Yan entered Peking University in 1995 and earned a bachelor's degree in 1999 and a PhD in applied mathematics in 2004 under Cheng Qiansheng (程乾生). He then did a postdoctorate at the Chinese University of Hong Kong's multimedia lab under Tang Xiaoou (汤晓鸥), the late pioneer of face recognition, followed by time at the University of Illinois at Urbana-Champaign with Thomas Huang (黄煦涛). In 2007 he joined the National University of Singapore (NUS), where he founded the Machine Learning and Computer Vision Lab.
He is a rare five-way fellow: the Singapore Academy of Engineering, AAAI, ACM, IEEE, and IAPR have all named him a fellow. That breadth reflects his range across computer vision, machine learning, and multimedia analysis.
The academia-industry loop
What sets Yan apart is how often he has stepped out of the university and back. In 2015 he joined 360 Group (三六零) as vice president, chief scientist, and head of its AI research institute. In 2019 he became chief technology officer of Yitu (依图), a Chinese computer-vision unicorn. In 2021 he moved to Southeast Asian internet group Sea (Sea Limited), where he served as chief scientist and founded Sea AI Lab (SAIL) in Singapore.
In September 2023 he joined Kunlun Tech (昆仑万维) as co-CEO of its TianGong AI (天工智能) unit and the first head of its 2050 Global Research Institute (2050全球研究院), building research centers in Singapore, London, and Silicon Valley. Fourteen months later, in November 2024, Kunlun announced he would step down as institute head and become an honorary advisor, while also resigning the TianGong co-CEO role. In his own words at the time, the AIGC environment had shifted "from basic research to application deployment," and he felt he had "completed the assigned mission (已完成既定使命)."
That loop — build a lab, ship products, return to research — is now his signature. As of 2025 he is a professor at NUS and describes his current direction as "e-AGI": executable, deployable, and evolvable artificial general intelligence, built on open-source models.
The 2025 work: memory for talking machines
VoiceMem is the clearest statement of where Yan's attention is going. Today's voice assistants reset the moment a session closes; the next time you talk, they have forgotten your preferences. Yan's team — working with researchers from Nanyang Technological University, NUS, Tsinghua, the Chinese University of Hong Kong, and Open Interaction Lab — built a framework with two "brains": one organizes facts ("what happened"), the other manages personality, attitude, and emotion ("who I am talking to"). The design retrieves only a handful of memories yet scores well above earlier memory baselines on long-conversation benchmarks, while keeping latency low enough for live speech.
He has also pushed multimodal modeling. In October 2025 his team released PaDT, a multimodal large model that outputs visual tokens directly instead of guessing coordinates — an approach aimed at finer visual grounding. Both projects reflect a consistent theme: make models not just bigger, but more efficient, more deployable, and better at interacting in real time.
Why his path matters
Yan's career is a window into how Chinese AI talent circulates. The same person who advised governments and founded corporate labs also keeps publishing and mentoring students. For a field obsessed with model size, his emphasis on efficiency and deployment is a counterweight — and a reminder that the hardest problems are often integration and memory, not raw parameters.
What readers can do now
- If you build voice or chat products, study open-source memory frameworks like VoiceMem before bolting on a vector database — latency and personality continuity are separable problems.
- If you are evaluating an AI scientist's claims, look at reproducibility: Yan's group tends to ship code, which is a healthier signal than citation counts alone.
- For students choosing a path, his loop shows that industry stints and academic return are not mutually exclusive; each sharpens the other.
Honest limitations
This article draws on Yan's published record, company announcements (Kunlun's official November 2024 statement), and secondary reporting on the 2025 releases. The specific benchmark numbers for VoiceMem and PaDT come from project write-ups and Chinese tech media rather than peer-reviewed papers I could independently re-run, so I have described their direction rather than vouching for exact scores. His citation totals are widely cited but vary by source, so I have avoided stating a precise figure. Some details of his early career timeline differ slightly across profiles.
Key takeaways
- Shuicheng Yan is a five-way fellow whose Network In Network helped reshape modern neural-network design.
- His career repeatedly moves between top universities and major Chinese tech firms, most recently Kunlun Tech, where he built the 2050 Global Research Institute before returning to NUS.
- In 2025 his focus turned to making voice agents remember — a deployable, efficiency-first view of where practical AI is heading.