Right for the Right Concept: Revising Neuro-Symbolic Concepts by Interacting with their Explanations

Most explanation methods in deep learning map importance estimates for a\nmodel's prediction back to the original input space. These "visual"\nexplanations are often insufficient, as the model's actual concept remains\nelusive. Moreover, without insights into the model's semantic concept, it is\ndifficult -- if not impossible -- to intervene on the model's behavior via its\nexplanations, called Explanatory Interactive Learning. Consequently, we propose\nto intervene on a Neuro-Symbolic scene representation, which allows one to\nrevise the model on the semantic level, e.g. "never focus on the color to make\nyour decision". We compiled a novel confounded visual scene data set, the\nCLEVR-Hans data set, capturing complex compositions of different objects. The\nresults of our experiments on CLEVR-Hans demonstrate that our semantic\nexplanations, i.e. compositional explanations at a per-object level, can\nidentify confounders that are not identifiable using "visual" explanations\nonly. More importantly, feedback on this semantic level makes it possible to\nrevise the model from focusing on these factors.\n