Integrated gradients
Integrated gradients: A method assigning prediction attribution by integrating gradients along a path from a baseline. It computes contribution between a reference and actual input, satisfying completeness. Baseline choice is part of the interpretation. Examples include Black baseline for images, Reference token for text, Attribution in differentiable models.
In italiano: Integrated GradientsLegge V · Overfitting, bias e collasso dei modelliIt computes contribution between a reference and actual input, satisfying completeness. Baseline choice is part of the interpretation.
Formula
IGᵢ(x) = (xᵢ−x′ᵢ) ∫₀¹ ∂F(x′+α(x−x′))/∂xᵢ dα
Examples
- Black baseline for images
- Reference token for text
- Attribution in differentiable models