Fault tolerance techniques are generally based around a common concept that is redundancy whose measurement is required. A suite of four semantic metrics is proposed to assess program redundancy and reflect their ability to tolerate faults. Literature shows that one of these metrics, namely state redundancy, is limited to compute program redundancy only in their initial and final states and ignores their internal states. Consequently, the authors focus in this paper to overcome this shortcoming by proposing a new redundancy-based semantic metric that computes the redundancy of the different program states including internal ones. The empirical study they perform shows that the proposed metric is a measure of program redundancy in one side and is an error detection indicator in another side. Moreover, they demonstrate that it is more accurate than the basic state redundancy metric in detecting masked errors. It is useful for testers to indicate if a tested program is error-free and to pinpoint the presence of masked errors even if the final states are equal to the expected ones.
Comparison of state-of-the-art deep learning APIs for image multi-label classification using semantic metrics
