Enriching Visual with Verbal Explanations for Relational Concepts -- Combining LIME with Aleph

With the increasing number of deep learning applications, there is a growing\ndemand for explanations. Visual explanations provide information about which\nparts of an image are relevant for a classifier's decision. However,\nhighlighting of image parts (e.g., an eye) cannot capture the relevance of a\nspecific feature value for a class (e.g., that the eye is wide open).\nFurthermore, highlighting cannot convey whether the classification depends on\nthe mere presence of parts or on a specific spatial relation between them.\nConsequently, we present an approach that is capable of explaining a\nclassifier's decision in terms of logic rules obtained by the Inductive Logic\nProgramming system Aleph. The examples and the background knowledge needed for\nAleph are based on the explanation generation method LIME. We demonstrate our\napproach with images of a blocksworld domain. First, we show that our approach\nis capable of identifying a single relation as important explanatory construct.\nAfterwards, we present the more complex relational concept of towers. Finally,\nwe show how the generated relational rules can be explicitly related with the\ninput image, resulting in richer explanations.\n