AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

XAI for Transformers: Better Explanations through Conservative Propagation

Ameen Ali, Thomas Schnake, Oliver Eberle, Grégoire Montavon, Klaus-Robert Müller, Lior Wolf

ICML 2022 · Architectures and efficiency › Attention and transformers · approach: Transformer block and depth design · 25 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Add LRP-style detaching of attention gating and LayerNorm rescaling (AH+LN) over plain Gradient x InputTransformer and Graphormer explanations · Perturbation activation/pruning tasks across 9 datasetsAUAC / AU-MSE: e.g. SST-2 AUAC .847 -> .908; SST-2 AU-MSE 2.10 -> 1.56Improved
Detach attention gating terms only (AH rule) (ablation)Gradient x Input on Transformer · SST-2AUAC: .847 -> .855Improved
Detach LayerNorm rescaling only (LN rule) (ablation)Gradient x Input on Transformer · SST-2AUAC: .847 -> .907Improved
Detach LayerNorm rescaling only (LN rule) (ablation)Gradient x Input on Transformer · SST-2AU-MSE: 2.10 -> 1.78Improved
Detach attention gating terms only (AH rule)Gradient x Input on Transformer · MNIST superpixels (Graphormer)AU-MSE: 18.06 -> 18.03No clear effect
Detach attention gating terms only (AH rule)Gradient x Input on BACE Graphormer · BACEAUAC: .646 -> .645No clear effect
Detach attention gating terms only (AH rule)Gradient x Input on BACE Graphormer · BACEAU-MSE: 2.06 -> 2.08Hurt
Detach attention gating terms only (AH rule)Gradient x Input on IMDB · IMDBAUAC: .920 -> .911Hurt
Detach attention gating terms only (AH rule)Gradient x Input on IMDB · IMDBAU-MSE: 0.87 -> 0.77Improved
Detach LayerNorm rescaling only (LN rule) vs GI on BACEGradient x Input on BACE Graphormer · BACEAUAC: .646 -> .702Improved
Gradient x Input baseline vs raw attention and GAE baselinesExplanation methods · Perturbation tasks across datasetsAUAC: e.g. IMDB GI .920 vs Rollout .738Improved
Use Gradient x Input with unmodified gradients (no conservation fix) in LayerNorm/attentionGradient x Input on Transformer · SST-2 conservation checkconservation (sum of relevance vs output)Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy