Applying an Axiomatic Model Attribution Technique to Deep-Learning Image Modeling

Integrated Directional Gradients (IDG) is a novel technique for producing high level explanations
for Natural Language Processing models. In this paper, we extend this method to Deep Neural
Networks for image processing. In recent years, these black-box models have become incredibly
effective at identifying similarities between large numbers of input features. Our new methodology
produces attribution scores for groups of input features that specify how influential they are on the
model’s output. In image processing, we can generate these feature groups with image segmentation
to further understand the output of an image classification model. We train two state-of-the-art
models for both image classification and image segmentation on benchmark datasets to help us
evaluate which image features affect the model’s performance.