More about HKUST
Towards Automated and Trustworthy Machine Learning
Speaker: Dr. Minhao Cheng University of California, Los Angeles Title: "Towards Automated and Trustworthy Machine Learning" Date: Tuesday, 20 April 2021 Time: 10:00 am - 11:00 am Zoom Link: https://hkust.zoom.us/j/465698645?pwd=c2E4VTE3b2lEYnBXcyt4VXJITXRIdz09 Meeting ID: 465 698 645 Passcode: 20202021 Abstract: Deep neural networks have achieved unprecedented success over a variety of tasks and across different domains. At the same time, it has been shown that DNNs models are vulnerable to a very small human-imperceptible perturbation. In this talk, I will first introduce how to develop a query-efficient framework to generate such perturbation which could apply to industrial-strength image classifiers in real-world scenarios. Furthermore, I will show a single query oracle for retrieving signs of directional derivative could be utilized to largely improve the query efficiency. Lastly, to develop an accurate and trustworthy model, I will talk about how to automate the manual process of architecture design in a differentiable manner and improve its limitation on the architecture selection. *************************** Biography: Minhao Cheng obtained his Ph.D. degree in the Department of Computer Science from the University of California, Los Angeles under the supervision of Prof. Cho-Jui Hsieh. His research focus is broadly on machine learning with a focus on machine learning robustness and AutoML. He has published over 15 papers on top-tier AI conferences including ICML, NeurIPS, ICLR, ACL, AAAI, etc. He is a recipient of the ICLR 2021 Outstanding Paper Award.