Structural Coverage Criteria for Neural Networks Could Be MisleadingNIER
There is a dramatically increasing interest in the quality assurance for DNN-based systems in the software engineering community. An emerging hot topic in this direction is structural coverage criteria for testing neural networks, which are inspired by coverage metrics used in conventional software testing. In this short paper, we argue that these criteria could be misleading because of the fundamental differences between neural networks and human written programs. Our preliminary exploration shows that (1) adversarial examples are pervasively distributed in the finely divided space defined by such coverage criteria, while available natural samples are very sparse, and as a consequence, (2) previously reported fault-detection “capabilities” conjectured from high coverage testing are more likely due to the adversary-oriented search but not the real “high” coverage.
Fri 31 MayDisplayed time zone: Eastern Time (US & Canada) change
14:00 - 15:30 | Testing of AI SystemsNew Ideas and Emerging Results / Demonstrations / Technical Track at Place du Canada Chair(s): Marija Mikic Google | ||
14:00 20mTalk | CRADLE: Cross-Backend Validation to Detect and Localize Bugs in Deep Learning LibrariesTechnical Track Technical Track Hung Viet Pham University of Waterloo, Thibaud Lutellier , Weizhen Qi University of Science and Technology of China, Lin Tan Purdue University Pre-print | ||
14:20 20mTalk | Guiding Deep Learning System Testing using Surprise AdequacyTechnical Track Technical Track Jinhan Kim KAIST, Robert Feldt Chalmers University of Technology, Shin Yoo Korea Advanced Institute of Science and Technology Authorizer link Pre-print | ||
14:40 20mTalk | DeepConcolic: Testing and Debugging Deep Neural NetworksDemos Demonstrations Youcheng Sun University of Oxford, Xiaowei Huang University of Liverpool, Daniel Kroening University of Oxford, James Sharp Defence Science and Technology Laboratory (Dstl), Matthew Hill Defence Science and Technology Laboratory (Dstl), Rob Ashmore Defence Science and Technology Laboratory (Dstl) | ||
15:00 10mTalk | Towards Improved Testing For Deep LearningNIER New Ideas and Emerging Results Pre-print | ||
15:10 10mTalk | Structural Coverage Criteria for Neural Networks Could Be MisleadingNIER New Ideas and Emerging Results Zenan Li Nanjing University, Xiaoxing Ma Nanjing University, Chang Xu Nanjing University, Chun Cao Nanjing University Pre-print | ||
15:20 10mTalk | Robustness of Neural Networks: A Probabilistic and Practical PerspectiveNIER New Ideas and Emerging Results |