Recently the program committee of ESEC/FSE'20 annouced the list of papers that received ACM SIGSOFT Distinguished Paper Awards at the conference. One of them is "Detecting Numerical Bugs in Neural Network Architectures", which was led by researchers at the Programming Languages Lab. The award recognizes the top 10% of the accepted papers at the conference.

The paper makes the first attempt to conduct static analysis for detecting numerical bugs in deep learning at the architecture level. Detecting bugs at the architecture level provides additional benefits that detecting bugs at the model level does not provide.  The experiment results show that proposed approach out-performs other tensor and numerical abstraction techniques on accuracy without losing scalability. On real-world architectures, it reports 529 warnings within 2.6-135.4 seconds per architecture, where 299 warnings are true positives.

The team led by Prof. Junjie Chen from Tianjin Unviersity, an alumnus of the lab, also published  a paper at the conference. Prof. Chen was co-advised by Prof. Yingfei Xiong and graduated last year.