Integrated Gradients
XAI method that computes feature attributions by integrating gradients along a path from a baseline to the actual input.
Integrated Gradients computes axiomatically correct feature attributions for deep learning – the most theoretically grounded gradient-based XAI approach.
Explanation
Integrated Gradients satisfies two important axioms: Sensitivity (if a feature changes output, it receives attribution) and Implementation Invariance (same function = same attribution). This makes it theoretically more robust than vanilla gradients or DeepLIFT.
Marketing Relevance
Standard attribution for deep learning in production. Implemented by Google in Cloud AI Explanations and Captum (PyTorch).
Common Pitfalls
Baseline choice strongly influences results. Path dependency with non-linear models. More compute-intensive than simple gradients.
Origin & History
Sundararajan, Taly & Yan published Integrated Gradients in 2017 (ICML). Google implemented it in Cloud AI Explanations. Meta integrated it in Captum. The method became the standard for deep learning attribution.
Comparisons & Differences
Integrated Gradients vs. SHAP (DeepSHAP)
Integrated Gradients uses path integration (axiomatic); DeepSHAP uses Shapley approximation (faster but less exact for deep networks).
Integrated Gradients vs. Saliency Map
Saliency maps use one gradient step (noisy); Integrated Gradients accumulates over the entire path (more robust).