ConEx: Human-Interpretable Saliency Maps via Concept-Aware Attribution

ConEx. Human-Interpretable Saliency Maps via Concept-Aware Attribution. By Yehonatan Elisha and three others, from Tel Aviv University and the Open University. Published on the third of October. Its code repository is public but slight: seven source files.

Many visual explanation methods in computer vision show which pixels mattered to a prediction. What they struggle to do is connect those low-level cues to concepts a person would recognise as meaningful. A map of important pixels that cannot be tied to anything semantic limits how far the explanation can be interpreted, and how far it can be trusted.

Concept-based Explanations, or ConEx, builds a bridge between two things: saliency visualization and concept-based reasoning. The aim is to have both at once, explanations that are faithful to the model and interpretable to a person. ConEx discovers class-specific concepts automatically and represents each one as a concept activation vector, a CAV.

Those vectors are learned without manual supervision. Concept discovery relies on segmentation masks, and the masks introduce noise, so ConEx uses a masking mechanism specific to the architecture being explained to reduce that noise and make the concepts purer. From there it generates saliency maps that show where each concept appears in the image and how that concept contributes to the prediction. Each highlighted region therefore comes with two answers rather than one: which concept is present there, and what that concept did to the model's output.

To judge whether the learned concepts are reliable, the authors propose two complementary metrics. Vector-Concept Match and Concept-Class Match quantify how well concepts align, and allow direct comparison with existing methods. Across extensive experiments in diverse settings, ConEx is reported to achieve state-of-the-art performance on benchmarks for faithfulness, for segmentation, and for concept quality.

The shift is in what a saliency map is a map of. Instead of answering only where the model looked, ConEx tries to answer which concept it found there and how much that concept counted, so that a heat map arrives already labelled.