Randomly dropping pixels during gradient integration yields explanations with higher measured input-explanation mutual information than noise-based or linear sampling on ImageNet pretrained models.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CV 1years
2025 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
Sampling Matters in Explanations: Towards Trustworthy Attribution Analysis Building Block in Visual Models through Maximizing Explanation Certainty
Randomly dropping pixels during gradient integration yields explanations with higher measured input-explanation mutual information than noise-based or linear sampling on ImageNet pretrained models.