Applied Mathematics Seminar——Specification-Driven Post-Deployment Repair of Neural Networks
报告人：Wenchao Li（Boston University）
Deep neural networks, despite their impressive performances on a wide range of problems, are not immune to making mistakes, e.g., misclassifying an image or producing the wrong control action. In this talk, I will first provide a taxonomy of post-deployment neural network repair techniques. Then I will describe REASSURE, the first sound and complete repair technique for ReLU networks with strong locality and minimality guarantees. The key idea of REASSURE is to leverage the continuous piecewise linear property of ReLU networks to synthesize a patch network that is tailored to the linear region where the buggy input resides, which when combined with the original network, provably corrects the behavior on the buggy input. Experimental results show that REASSURE outperforms competing techniques by several orders of magnitude across a number of benchmarks and metrics. I will also present REGLO, a novel technique that enables provable repair of neural networks for global robustness properties. I will conclude the talk by touching upon the other facets of building safe and trustworthy learning-enabled systems and highlighting some of our other recent efforts in this community-wide endeavor.
Wenchao Li is an Assistant Professor in the Department of Electrical and Computer Engineering at Boston University with affiliate appointments in Computer Science and System Engineering. Prior to joining BU, he was a Computer Scientist at SRI International, Menlo Park. He received his Ph.D. in Electrical Engineering and Computer Sciences from the University of California, Berkeley in 2013. His research interests lie broadly in dependable computing with a focus on exploring the new fundamental science at the intersection of formal methods and machine learning. He received the ACM Outstanding Ph.D. Dissertation Award in Electronic Design Automation for his work on specification mining and the Leon O. Chua Award for outstanding achievement in nonlinear science.
加入 Zoom 会议