DeepLIFT
Algorithms AIMech:DeepLIFT DRAFT
Aliases: Deep Learning Important FeaTures
A neural-network attribution method that distributes an output's difference from a reference activation to input differences using backward contribution propagation.
Identity Notes & Concept Scope
This record covers the 2017 DeepLIFT method including distinct local rules; it is not identical to gradients, integrated gradients, Deep SHAP or the earlier Rescale-only formulation.
Claims & Grounded Evidence
1 Claim
Mechanism
#attribution-method
Reference inputs determine reference activations. Local contribution multipliers compose backward subject to summation-to-delta; Linear, Rescale and RevealCancel rules determine local allocation, with RevealCancel treating positive and negative differences separately.
⚠️ Curated Limitations & Negative Results
- Reference and local-rule choices affect attribution. The paper discusses cases favoring Rescale or RevealCancel rather than one universally superior rule.
- Contribution conservation is not a proof of causal feature importance or exact Shapley values for an arbitrary network.
Artificial Intelligence Ontology (AIO) Grounding
Mapped AIO Classes
| AIO Target Class | Mapping Relation | Mapping Rationale |
|---|---|---|
AIO:ArtificialNeuralNetwork
Artificial Neural Network
|
related_match | The defining method propagates contributions through an artificial neural network. The related mapping identifies that direct analysis object, not architecture identity or subclasshood. |
Evidence Sources & Primary Literature
1 Source| Source ID | Publication / Document | Type / Access | Reference & Checksum |
|---|---|---|---|
original-paper
|
Avanti Shrikumar, Peyton Greenside, Anshul Kundaje
|
primary paper
full_text
|
sha256:bfdc83443542…
|
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/DeepLIFT.yaml