Explainability

Explanation-based Weakly-supervised Learning of Visual Relations with Graph Networks

A novel weakly-supervised method for visual relationship detection that relies on minimal image-level predicate labels, graph networks, and attribution-based explanations.

Explainability Techniques for Graph Convolutional Networks

How do Sensitivity Analysis, Guided Backpropagation, and Layer-Wise Relevance Propagation behave when applied to Graph Neural Networks?