What to Blame? On the Granularity of Fault Localization for Deep Neural Networks

Abstract

Validating Deep Neural Networks (DNNs) used for classification is of paramount importance; an approach for this consists in (i) executing the DNN over the test dataset, (ii) collecting information about classifications, and (iii) applying fault localization (FL) techniques to identify the neurons responsible for the misclassifications. DNNs can have multiple misclassification types, and so neurons responsible for one type could be different from those responsible for another type. However, depending on the granularity of the analyzed dataset, FL may not reveal these differences: failure types more frequent in the dataset may mask less frequent ones. We here propose a way to perform FL for DNNs that avoids this masking effect by selecting test data in a granular way. We conduct an empirical study, using a spectrum-based FL approach for DNNs, to assess how FL results change by changing the granularity of the analyzed test data. Namely, we perform FL by using test data with two different granularities: following a state-of-the-art approach that considers all misclassifications for a given class together, and the proposed fine-grained approach. Results show that FL should be done for each misclassification, such that practitioners have a more detailed analysis of the DNN faults and can make a more informed decision on what to fix in the DNN.