AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

On Spectral Properties of Gradient-based Explanation Methods

Amir Mehrpanah, Erik Englesson, Hossein Azizpour

ECCV 2024 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 1 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
use squared gradients instead of gradients in SmoothGrad (ablation)SmoothGrad · ImageNetInsertion: 0.212 -> 0.465Improved
use squared gradients instead of gradients in Integrated Gradients (ablation)IG · ImageNetInsertion: 0.201 -> 0.340Improved
use squared gradients instead of gradients in SmoothGrad (ablation)SmoothGrad · Food101Deletion: 0.137 -> 0.162Hurt
use squared gradients instead of gradients in XIG (ablation)XIG · Food101Deletion: 0.177 -> 0.263Hurt
use per-image optimal perturbation scale (max cosine similarity) in SG^2 (ablation)SG^2 · ImageNetInsertion: 0.465 -> 0.491Improved
use per-image optimal perturbation scale in SG^2 (ablation)SG^2 · Food101Insertion: 0.371 -> 0.394Improved
use per-image optimal perturbation scale in SG^2 (ablation)SG^2 · ImageNetDeletion: 0.157 -> 0.187Hurt
aggregate over noise scales with uniform explanation prior (SpectralLens) instead of Dirac prior (ablation)SG^2 · ImageNetInsertion: 0.465 -> 0.493Improved
aggregate over noise scales with uniform explanation prior (SpectralLens) (ablation)SG^2 · Food101Deletion: 0.162 -> 0.157Improved
apply input multiplication (XIG) instead of IG (ablation)Integrated Gradients · Food101Insertion: 0.145 -> 0.101Hurt

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