Selected Impacts (IEEE AI 10 to Watch, MIT TR35 China, IBM PhD Fellow, World’s Top 2% Scientists, IJCAI Early Career Spotlight, 10+ best paper awards, 10+ patents)
At Sony AI, I built a global Privacy-preserving Machine Learning (PPML) and Vision Foundation Model (VFM) team from 0-1. Our team had supported multiple business units such as Sony Semiconductor Solutions (SSS) --#1 in global image sensor market, on the development and PoC of a series of novel on-device AI solutions that can support multi-task training under semi/un-supervised setting. We have deployed many AI models in various environment (cloud, consumer and edge devices even with few MB) to simultaneously conduct multiple CV tasks (classification, object detection, semantic segmentation, pose estimation, etc) in an accurate, private and fast manner. Our team delivered a VFM V1 in 9 months with 12 tasks, VFM V1.5 in 6 months with 17 tasks, a Stable Diffusion v1/v2-quality model trained from scratch in just 2.5 days using 8 H100 GPUs. The outcomes of both projects were accepted by CVPR’25 (top 1 computer vision conference). Our VFM is by far the most comprehensive and well-performing VFM of million scale, our MicroDIT model is by far the cheapest image generation model (1890 USD). We’re currently developing low-cost and powerful VLM (understanding tasks like captioning, VQA, reasoning) and more unified foundation model (supporting perception, understanding and generation). Our team works cross-functionally with engineering and product teams.
Before Sony AI, when I was a team leader in Ant Group. My team proposed a series of new algorithms (Fedsplit-GNN, Fed-KG) for PoC and landing of collaborative and privacy-preserving risk modelling in Fintech. Our team successfully realized the first landing of privacy-preserving multi-party risk governance in security and risk management business group (1k+ people) in Ant Group (Press). It had largely helped improve the fraud detection rate through the federation between Alipay (world’s largest mobile payment platform with billion++ users) and other business units across Ant Group and Alibaba.
Before Ant Group, I worked at ANU as a Research Fellow (Level B3, equivalent to lecturer/assistant professor). Back to late 2017 and early 2018 when I was at IBM, I was the first person who proposed the concept of Collaborative Fairness in Federated Learning. Since then, Collaborative Fairness has attracted substantial attention from both academia (KDD'21 tutorial, CIKM, AAAI, etc) and industry (appeared in the white paper of China Academy of Information and Communications Technology (CAICT), landed in multiple companies' products, etc).
Sponsorship and Press
My past research projects were supported by an IBM Ph.D. Fellowship, ANU Translational Fellowship, Australian Research Council (ARC), EU FP7, National Science Foundation, etc. My past research has been featured or quoted by Nature, IEEE, MIT Technology Review, TechCrunch, Wired, Bloomberg, Marktechpost, Towards Data Science (679K+ follower by 2023), SyncedReview, IBM, Microsoft, RE•WORK, insideainews, Toutiao, AI Technology Review, Paperweekly, AI era, CSDN, NetEase Finance, Sina Weibo, Zhuanzhi, iresearch, etc.