AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Beyond Single Path Integrated Gradients for Reliable Input Attribution via Randomized Path Sampling

Giyoung Jeon, Haedong Jeong, Jaesik Choi

ICCV 2023 · Label-efficient and robust learning › Recognition and long-tail · approach: Feature attribution methods · 4 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
aggregate attributions over multiple stick-breaking sampled paths instead of a single straight pathIntegrated Gradients · ImageNet val, VGG-16insertion: 0.096 -> 0.443Improved
aggregate attributions over sampled pathsIntegrated Gradients · ImageNet val, VGG-16deletion: 0.036 -> 0.019Improved
SPI versus IG under ROAR with 10% pixels removedIntegrated Gradients · CIFAR-10 ROAR ResNet-18test accuracy: 39.70% -> 31.00%Improved
sample paths with stick-breaking process with uniform base distribution instead of hypergeometric distribution (ablation)SPI path sampling · path distribution analysisrepresentation power of sampled pathsImproved
use concentration alpha = 10 rather than low alpha (ablation)SPI · ImageNet insertion/deletioninsertion/deletion gameImproved
FullGrad versus other baselines on ROAR at low removalFullGrad · CIFAR-10 ROAR 10% removedtest accuracy: 67.49% at 10% removed vs IG 39.70%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