Explaining the inner workings of deep neural network models have received considerable attention in recent years. Researchers have attempted to provide human parseable explanations justifying why a model performed a specific classification. Although many of these toolkits are available for use, it is unclear which style of explanation is preferred by end-users, thereby demanding investigation. We performed a cross-analysis Amazon Mechanical Turk study comparing the popular state-of-the-art explanation methods to empirically determine which are better in explaining model decisions. The participants were asked to compare explanation methods across applications spanning image, text, audio, and sensory domains. Among the surveyed methods, explanation-by-example was preferred in all domains except text sentiment classification, where LIME's method of annotating input text was preferred. We highlight qualitative aspects of employing the studied explainability methods and conclude with implications for researchers and engineers that seek to incorporate explanations into user-facing deployments.

Weaknesses: The term'unified' should be revised as the paper addresses a partial unification. For instance, the unified framework does not take into account a closed loop between the DNN and the explanation method (the explanation method can be itself another DNN interacting in a double sense with the prediction DNN) or other two-stage adaptive networks [1], [2]. In addition, an alternative to example based explanation is'opening the black box' in terms of intra-layer and inter-layer statistical properties of a DNN [3]: these may be enough to explain lack of generality (and thus absence of recommendation) of a given network depending on the input available data and the classification paradigm considered. Thus, a positioning must be provided with respect to the above issues in order to make the paper more informative with respect to the literature. The weak spots of the analysis are twofold.

The paper is an empirical study on the types of explanations preferred by users using AMT. All reviewers found that problem was important and found the study interesting. However, one reviewer argued that while this was a good first step, it does not address the fact evaluating explanations is an ill-posed problem. Three reviewers found that this study is interesting enough to the NeurISP community even as a first step.

Explaining the inner workings of deep neural network models have received considerable attention in recent years. Researchers have attempted to provide human parseable explanations justifying why a model performed a specific classification. Although many of these toolkits are available for use, it is unclear which style of explanation is preferred by end-users, thereby demanding investigation. We performed a cross-analysis Amazon Mechanical Turk study comparing the popular state-of-the-art explanation methods to empirically determine which are better in explaining model decisions. The participants were asked to compare explanation methods across applications spanning image, text, audio, and sensory domains.