Mukund Sundararajan and colleagues, 2017

“We identify two fundamental axioms—Sensitivity and Implementation Invariance—that attribution methods ought to satisfy.”

Mukund Sundararajan and colleagues wrote this while addressing designing neural-network attributions with clear principles rather than visually appealing heuristics. The research integrated input gradients along a path from a baseline to the observed example. In the original integrated gradients setting, the claim is bounded by reported data and experimental conditions. That specificity matters because the selected baseline and path materially shape the attribution and may not reflect a meaningful real-world counterfactual.

For integrated gradients, the quotation matters because the method introduced a simple attribution method grounded in explicit axioms. By explaining that the work integrated input gradients along a path from a baseline to the observed example, the authors made their design testable. Later studies can compare alternatives for designing neural-network attributions with clear principles rather than visually appealing heuristics and ask whether the evidence transfers to evaluating feature importance while documenting baseline choices and comparing multiple explanation methods.

Applied today, integrated gradients remains useful for evaluating feature importance while documenting baseline choices and comparing multiple explanation methods. Responsible teams should define intended users, a baseline for designing neural-network attributions with clear principles rather than visually appealing heuristics, and a rollback signal before deployment. Because the selected baseline and path materially shape the attribution and may not reflect a meaningful real-world counterfactual, evaluation also needs rare-case tests, ongoing monitoring, an appeal path for affected people, and named ownership for correction.

The central caution is that the selected baseline and path materially shape the attribution and may not reflect a meaningful real-world counterfactual. For integrated gradients, the result supports the finding that the method introduced a simple attribution method grounded in explicit axioms, but not every possible use. Deployment should include independent review, incident reporting, and a practical way to withdraw the system. The sound reading preserves the contribution while leaving room to revise it when real-world evidence disagrees. The integrated gradients quotation endures because integrated input gradients along a path from a baseline to the observed example offers a testable response to designing neural-network attributions with clear principles rather than visually appealing heuristics. Readers considering evaluating feature importance while documenting baseline choices and comparing multiple explanation methods should reproduce the result, publish negative cases, and compare simpler baselines. They should also explain that the selected baseline and path materially shape the attribution and may not reflect a meaningful real-world counterfactual, assign responsibility, and prepare a correction path. Continued value for integrated gradients depends on transparent evidence, careful limits, and correction when new conditions expose failure. That balance lets later work build on the contribution without erasing the limits that make its evidence interpretable.