My research focuses on the intersections between hardware, algorithms, and security in the Machine Learning world. I am generally interested in making Machine Learning systems fast, reliable and efficient. Some of my research topics include:
Automatic Co-designing AI Systems with MASE: MASE aims to provide a unified representation for software-defined ML heterogeneous system exploration. My team expects to explore new cross-stack acceleration opportunities, from re-designing the algorithm to renovating the hardware architecture, to make AI workloads run faster.
Beyond Structure Data: I am interested in working on projects that involve unstructured and multimodal data, such as graphs, hypergraphs and combinational complex. We envision these data types would be the eanbler for the next generation of AI systems that goes beyond simple images and text.
Efficient AI: I am interested in different algorithmic aspects of efficient AI, including efficient training, efficient inference, efficient model search and efficient model deployment with state-of-the-art GenAI models (eg. language and diffusion models).
System-level AI Safety: with the increasing capability of GenAI models and the growing complexity of AI systems, I am interested in projects that focus on the system-level AI safety, including robustness, security, and red-teaming these models to understand new vulnerabilities.
See advice for students (this includes internships, undergraduate or master projects, and PhD applicants), and information for collaborators, if you want to contact me for study or collaborations.
I lead the DeepWok Lab, and most of the research related updates are in the lab wiki.