Bio

Hi, my name is Wushi Dong (董悟时). I currently work as an Applied Scientist at AWS, specializing in optimizing Deep Learning inference efficiency through compiler and HPC techniques. I am deeply committed to maximizing computing resources’ efficiency and harnessing the full potential of deep learning technologies to capitalize on real-world opportunities at scale. With a background in theoretical physics and expertise in both software and hardware, I bring a unique perspective to the intersection of technology and innovation.

I obtained my Ph.D. in Physics from the University of Chicago, advised by Professor Peter B. Littlewood. I received my bachelor degree in Physics from the School of the Gifted Young (SGY) at the Univeristy of Science and Technology of China (USTC).

My track record includes leading successful projects in both academic and industrial environments:

• Led a project enabling AWS customers to optimize their TensorFlow 2-based Deep Learning models using SageMaker Neo, seamlessly deploying them on SageMaker Inference. This initiative significantly expanded model coverage and achieved superior performance compared to standalone Deep Learning compilers, empowering our customers with cutting-edge solutions tailored to their specific needs.
• Led the project of distributed training for 3D instance segmentation of brain imaging data at Argonne Leadership Computing Facilities (ALCF). Successfully scaled ML training to more than 130,000 CPU cores on the Theta supercomputer.
• Achieved state-of-the-art accuracy on video action recognition by using self-developed Python software implementing brain-inspired deep learning algorithms. This work was done during an internship at IBM Research AI.
• Constructed a complete simulation pipeline including parallel C++ software using MPI to model quantum electron transport in 2D semiconductor devices.

Competencies

• Deep learning (TensorFlow, PyTorch, Keras, Horovod, distributed training and inference)
• High-Performance Computing (MPI, SLURM, profiling, numerical techniques)
• Software Engineering (Python, C/C++, Docker, Bazel, Vim)
• Computer vision (3D segmentation, video classification)
• Theoretical Physics (mathematical modeling, statistics, device physics)

Projects

• Scaling distributed training of Flood-Filling Networks (FFN) on HPC infrastructure for brain mapping
• Two-stream CNNs for video action recognition
• Swan: An open-source C++ software for nanoscale quantum electron transport simulations