Integrated gradients
Mathematical Methods AIMech:IntegratedGradients DRAFT
Aliases: IG
A feature-attribution method that integrates a prediction's input gradients along the straight path from a chosen baseline to the input.
Identity Notes & Concept Scope
Straight-path gradient integration differs from a gradient at one input, layer-wise contribution propagation and Shapley averaging across feature orderings.
Claims & Grounded Evidence
1 Claim
Mechanism
#attribution-method
Each feature receives its input-minus-baseline difference multiplied by the integral of that feature's partial derivative along the path; the paper approximates the integral with a finite gradient sum.
⚠️ Curated Limitations & Negative Results
- Attributions depend on the baseline; an arbitrary zero-score baseline need not give a meaningful absence of signal.
- Numerical integration is approximate. This record makes no universal causal-faithfulness or superiority claim.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
related_match | The method attributes outputs of artificial neural networks to their inputs in the defining paper. This direct analysis relationship does not make the method a neural-network architecture. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Mukund Sundararajan, Ankur Taly, Qiqi Yan
|
primary paper
full_text
|
sha256:b17d267e7b8f…
|
Curation Provenance
Curator: Codex
Method: ai_assisted
Model: GPT-6
Created: 2026-10-11T03:54:14Z
Updated: 2026-10-11T03:54:14Z
Source File:
data/curated/IntegratedGradients.yaml