Gradient-Based Attribution

Gradient-based attribution studies how changes to inputs, internal representations, model parameters, or training examples affect a model’s outputs. Common approaches include saliency maps, Integrated Gradients and its path-based variants, class activation maps, and gradient-based influence methods for tracing training data or identifying important parameters. Researchers apply them to images, text, graphs, time series, and large language models, evaluating explanations with localization, deletion, faithfulness, robustness, and known-ground-truth benchmarks. A central tension is that gradients are efficient and broadly applicable, but their results can depend strongly on baselines, architectures, training choices, and evaluation metrics, so predictive accuracy alone does not establish that an attribution is faithful.