Dr. Shawn (Xiangru) Lian

About

I just left Kuaishou Technology (SEHK:1024) , where I spent a few years leading multiple engineering and algorithm teams building the company's large scale AI platform. I also served as a Senior Staff Research Scientist at Kuaishou Seattle AI Lab . We built reliable, scalable, and distributed systems with cutting edge technologies, including large scale ranking/personalization engines, high performance distributed deep learning infrastructure, and model compression frameworks.

I received my PhD degree from University of Rochester in 2019 working with my advisor Dr. Ji Liu . We published important results on distributed machine learning, including the theoretical justification of asynchronous SGD (NeurIPS 2015 spotlight ), and the first decentralized SGD with linear speedup (NeurIPS 2017 oral ).

I got started with programming before I went to preschool. Comprehensive engineering knowledge together with deep understanding of algorithms help me keep teams on the right direction to solve challenging real world problems.

Curriculum Vitæ: Encrypted PDF .

Education

Ph.D. (2015-2019) in Computer Science, University of Rochester
B.S. (2011-2015) in Physics, University of Science and Technology of China

Projects

PERSIA Our advanced GPU-based large scale learning system for ad recommendation and CTR prediction tasks. PERSIA was launched in 2018, and open sourced in 2021. PERSIA supports models with up to 100 trillion parameters, and is by far the fastest public recommendation model training framework.
- Training Deep Learning-based recommender models of 100 trillion parameters over Google Cloud.
- PERSIA, the largest recommended training system in the history of open source by far.
- Story: 640x Faster GPU Based Learning System for Ad Recommendation.
- Story: Innovation, Balance, and Big Picture: The Speed of Kwai Commercialization.
(collaborate with DS3 Lab)
Bagua Our deep learning training acceleration framework to speed up large scale training tasks, including data loader optimization, advanced distributed training algorithm, network communication optimization, and more. Bagua is our solution to solve the training bottleneck at Kuaishou Technology (more than a million videos uploaded per hour). It is now open sourced.

(collaborate with DS3 Lab)
Hammer An automatic deep learning model compression tool by us. Useful for trimming down large models while keeping the model's accuracy as good as possible. Hammer has saved thousands of GPU cards and enabled hundreds of complex models to be successfully deployed in Kuaishou Technology.
DouZero We released a new game AI for DouDizhu . The corresponding research paper was accepted by ICML 2021.

(collaborate with DATA Lab )
cproxy A little handy tool I created to apply proxy transparently on individual processes.

News

IEEE Low-Power Computer Vision Challenge Our team won the 2nd prize. Congrats!
I received the 30 New Generation Digital Economy Talents (30 位新生代数字经济人才) award from World Internet Conference and Big Data Digest . Thanks!
How we reduced 95% of the communication cost in training deep neural networks.
IBM cuts AI speech recognition training time from a week to 11 hours using our AD-PSGD algorithm.
IBM achieves 10x performance improvement with our decentralized training algorithms.
Our research on staleness-aware ASGD is applied in the largest Java deep learning framework Deeplearning4j .