Integrated Gradients
Definition
Integrated gradients assign contributions of input features to a model prediction by integrating gradients along the straight path from a chosen baseline to the input.
The method computes gradients at intermediate points between reference and actual input. The result assigns attributions to features and satisfies the axioms Sensitivity and Implementation Invariance formulated by the authors.
Simply explained
How to proceed step by step from a chosen reference point to the real example and measure which inputs change the prediction.
Why it matters in practice
Relevant for the analysis of differentiable models, especially with text, image, and tabular data.
Example from business
A classification model marks which input fields contribute the most to a risk assessment.
Limits and pitfalls
- Confusing attribution with causality; choosing an inappropriate baseline; considering only one explanation; not checking unstable explanations.
Origin & history
Sundararajan, Taly, and Yan introduced Integrated Gradients in 2017 as an axiomatic attribution method.
Use in mid-sized companies
- Possible application: model debugging, plausibility checks, and supportive explanations for differentiable models.
Frequently asked questions
Does IG show causality?
No. Why is the baseline important? It defines the reference point for the attribution.
Sources
- Axiomatic Attribution for Deep Networks · Sundararajan et al. (2017), Axiomatic Attribution for Deep Networks
- Volltext · Sundararajan et al., Volltext
Source verification: 10./11.10.2026.
Editorially reviewed on: