Hi! This is Yizhi, an experienced software engineer and an aspiring computer architect.

It was after the 1-year undergraduate research experience, that I realized I was not simply attracted to deep learning topics. I did substantial amount of work in bringing up new models and training, but building up software systems was really my thing.

A brief break from the academia, my work at Qualcomm and TikTok made it even more evident: it is fun to design and build fast systems. The beauty of it lies in its composability - we can build a splendid mansion from a myriad of bricks. We compose large systems from simple elements, and we’re able to maximize its performance by understanding these elements.

Apart from building efficient softwares, I also want to design the underlying architecture where they run fast. This is my future goal.

Research

I am interested in Computer Systems and Accelerators that does our jobs extremely fast.

I believe that computer systems will also be increasingly heterogeneous, namely different pieces sewed together, to maximize performance under limitations of computation power, memory bandwidth, and energy consumption.

I am eager to build such Frankenstein systems.

Projects

I’ve been closely working with Prof. Marcelo Mattar in a project about improving experience replay by re-prioritizing important transitions in the buffer. Our ideas came from his recent work in neuroscience, and we’ve proudly submited our preprint here. I’ve also worked with Prof. Michael Yip in UCSD Advanced Robotics and Controls Lab, where I aimed to automated surgical suturing using deep reinforcement learning and physics simulation using position-based dynamics (PBD). Here are the porfolios for both projects:

• Teach your computer to learn what’s relevant
• Teach a robot arm to do surgery