News

Learning relational visual concepts in an interpretable fashion is a challenge for pure deep learning approaches. This paper proposes a framework as part of a solution to this problem by combining deep learning methods of object detection with inductive logic programming (ILP) to train a visual concept learner. Object detection provides the basis for object-centric representations of images that are then used to inductively learn a classification rule. The paper further presents algorithms to explain any classification of the trained model by visualizing what a nearby example of the opposite class in the decision space could look like. An example concept from the domain of Kandinsky-Patterns is used to explain the framework and to provide a first evaluation in terms of performance, interpretability, and data-efficiency.