FMitF: Track I: Scalable and Quantitative Verification for Neural Network Analysis and Design

点击翻译按钮获取中文摘要

Neural Networks (NNs) have been successful in many areas including computer vision, speech recognition, and natural language processing. However, due to the increasing adoption of NNs in safety-critical and socially sensitive domains such as self-driving cars, robotics, computer security, criminal justice, and medical diagnosis, there is a pressing need for developing verification techniques that can provide guarantees about dependability and safety of NN applications. Formal-verification techniques can provide guarantees of correctness; however, existing approaches are not effective in analyzing real-world NNs with large numbers of neurons and complicated model structures. This project sets a comprehensive research agenda focusing on a holistic formal-verification framework for NNs that will provide a systematic and principled approach for developing dependable and safe NNs. It is intended to benefit major machine-learning applications such as autonomous driving and contribute to the leadership of the United States in software engineering and artificial intelligence. The research findings are being widely disseminated through open-source software packages, publications in premier conferences and journals, tutorials at teaching workshops, as well as specialized K-12 programs for exposing the young generation to the frontiers of software verification and machine-learning research. The team of researchers working on this project are integrating methods from the classical computing fields such as software engineering, automated verification, and formal methods to address the unique research challenges in the dependability and safety of NN applications. Specific research directions include 1) novel symbolic quantitative analysis techniques that provide sound results for establishing dependability and safety of the state-of-the-art NN models; 2) a set of effective system-level optimizations for computation/memory efficient NN verification with sufficient cross-framework portability and high verification efficiency; 3) advanced neural architecture design and training support for exploring and developing neural network models with verifiable robustness. The success of this research agenda is intended to enable a more complete and efficient software stack for improving the scalability of NN verification techniques and the robustness of NN applications.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.