Regression Fault Detection and Mitigation in the Evolution of Deep Learning Systems

Deep Learning (DL) is widely used in many industrial domains. The evolution of DL systems may induce regression faults, which may lead to serious consequences. Thus, ensuring the quality of DL systems during their evolution is crucial. Although there are existing approaches for understanding regression faults in traditional software, they are not suitable for DL systems because DL systems lack explicit logic structures. To better prevent regression faults in DL systems, in this work, we propose a three-step approach. First, through an interview study, we aim to understand the concerns of DL developers. Then, we propose methods for detecting regression faults. Finally, we suggest ways to mitigate these faults by understanding their potential causes.