Abstract
Network interpretation in context of explainable AI continues to gather interest not only because of the need to explain algorithm decisions, but also because of potential improvements that can be made to network design. A large pool of research effects have been made including explanation by training sample representer points, exhaustive feature occlusion methods, locally learned interpretable models, sensitivity methods using network gradients, and relevance models using layer-wise back-propagation. It is however a constant challenge to interpret different network architectures or even different network function layers given the multiplicity of models, tools, rules and assumptions. In addition, there are challenges in producing good interpretable results; in particular, that of jointly improving both the sensitivity and relevancy of each attribute contribution within a network to the final network decision. A unified decomposition rule based on new propositions about negative features and majority contribution is proposed in this paper to address these challenges. Furthermore, quantitative measures are discussed to address performance of both sensitivity and relevancy of interpretation.