Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models

Input Image "a hat" "long, shaggy hair" "an intelligent expression" SALF-CBM SALF-CBM "a small, dainty dog" "a ball" "a pot" Figure 1. Concept maps generated by our SALF-CBM. Inspired by human visual interpretation, our method first decomposes input images into spatially-localized structures, associated with familiar concepts, independent of a specific task. Explainability of task-specific outputs is obtained by training a final task layer on-top of these maps.