Scholar
Youjie Li
Google Scholar ID: 9NujVeYAAAAJ
MLSyser | PyTorch Post-doc | UIUC Ph.D.
Citations & Impact
Citations
1,037
H-index
12
i10-index
12
Publications
19
Co-authors
0
Publications
Resume
Academic Achievements
- VLDB'22: Harmony – enables training massive DNN models on commodity servers by overcoming GPU memory limits.
- MLSys'22: BNS-GCN – efficient full-graph GCN training via partition-parallelism and random boundary sampling (co-first author).
- ICLR'22: PipeGCN – pipelined feature communication for efficient full-graph GCN training.
- MobiCom'21: Visage – enables timely analytics for drone imagery (co-first author; code deployed in Microsoft’s FarmBeats).
- HotOS'21: Advocates training large DNNs on commodity hardware for broader accessibility.
- MICRO'19: DeepStore – in-storage acceleration for intelligent queries.
- ISCA'19: iSwitch – in-switch computing to accelerate distributed reinforcement learning.
- NeurIPS'18: PipeSGD – decentralized pipelined SGD framework for distributed deep net training.
Background
- Currently a Research Scientist at ByteDance, founding and building project veScale from scratch to support 99% of internal training jobs.
- Research focuses on Distributed Machine Learning Systems, spanning Efficient ML, System Optimization, and Hardware Acceleration.
- Pioneer in pipelined data parallelism (NeurIPS'18) and in-network acceleration for distributed training using SmartNICs (MICRO'18) and programmable switches (ISCA'19).
- Recent work includes massive model training systems (VLDB'22) and massive graph training systems (MLSys'22).
- Independently established his research direction in distributed ML systems during his Ph.D. at UIUC, where few students had previously worked in this area.