Integrated Gradients Explainer
Integrated gradients is a feature-attribution method for differentiable models. It begins with a baseline input, such as a black image or neutral embedding, and follows a path from that baseline to the real input.
At many points along the path, the method measures the gradient of the selected model output. It integrates those gradients and multiplies by the input difference to assign a contribution to each feature.
The method was designed to satisfy properties including sensitivity and implementation invariance. The attributions also approximately add up to the difference between the model’s output at the input and at the baseline.
Results still depend strongly on the baseline, path, output choice, and numerical approximation. A poor baseline can create misleading comparisons, so practitioners should justify it and test multiple reasonable alternatives.
Integrated gradients attributes a model prediction by accumulating gradients along a path from a chosen baseline to the actual input.
Key terms include baseline, path, gradient, integral, attribution, sensitivity, implementation invariance, completeness and approximation.
The method moves from a reference baseline toward the actual input.
Gradients sampled along that route are accumulated into feature contributions.
Changing the baseline can change the meaning and size of attributions.
Several defensible baselines and convergence checks should be examined.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.