Here is a selected list of my tech projects, which involve a wide range of topics such as autonomous driving, distributed systems, machine learning infrastructure, deep learning, robotics and data science. They are mostly group projects that I have led or participated in at university research labs and tech companies.
Pylot - A Software Package for Autonomous Driving Research
- Affiliated Research Group: Berkeley DeepDrive (BDD), Real-Time Intelligent Secure Execution Lab (RISELab)
- Tech Frameworks: Python, C++, Docker, ROS, DeepGTAV, OpenAI Gym
DescriptionDOS is a real-time operating system aiming at improving execution efficiency and accuracy on applications such as self-driving cars and robots. (In progress)
Hyperplane - A Distributed Deep Learning Training Platform
Affiliated Company: Pinterest
Advisors: Dmitry Kislyuk, Michael Xu
DescriptionHyperplane is a distributed model training platform. It has a web interface for users to specify how they want to train the models - what model to train, what data to use, which hyperparameters to set, how many workers you want, etc. After one click, Hyperplane is able to automatically launch jobs on several machines; more specifically, it processes queries, schedules training jobs, allocates GPU resources, and executes model training scripts in containers. (Code/Docs for this project is non-disclosable)
IDK Cascade - A Low-Latency Deep Learning Framework
Affiliated Company: Real-Time Intelligent Secure Execution Lab (RISELab)
Advisor: Prof. Joseph Gonzalez
- Tech Frameworks: Tensorflow
DescriptionIDK Cascade is an ensemble Convolutional Neural Network (CNN) structure. The motivation is that in most datasets there's distinction between easy and hard examples, so we want to use fast CNN to deal with easy examples and accurate CNN to deal with the hard ones so as to speed up the prediction process. We stack multiple CNNs together in an order of prediction latency, and train the whole structure end-to-end to make the model learn to tell apart easy and hard examples. IDK Cascade achieves 1.6x speedups in image classification tasks while maintaining the same accuracy; it also achieves nearly perfect accuracy (> 95%) in self-driving car tasks. Click below for our published paper.