Literature Review on Testing Deep Learning Based Systems

Abstract

Artificial Intelligence is continuing to increase in prevalence in today’s society. Advancements of artificial intelligence technology and the increasing attention and interest has brought upon a demand for a more sophisticated and precise system. In response, deep learning (DL) based systems have proven to be effective for various applications. Of the many applications, several systems involve safety critical characteristics which in turn calls for methods of verifying and assuring quality and safety of these systems. Testing of DL systems is a crucial aspect and may bring critical consequences if not done properly. However, DL testing differs greatly from traditional software testing which brings new challenges to the testing community. Recent research has presented several methods to solve these new problems, and many proved to be effective. We present a secondary study of the recent primary studies of the research area. Primary studies of DL testing have been tediously collected to provide an analysis of the recent trends of research in the area. Our study also identifies open challenges and suggestions for future works in the research area.