Reproduce “Explaining Convolutional Neural Networks using Softmax Gradient Layer-wise Relevance Propagation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Brian Kenji Iwana, Ryohei Kuroki, Seiichi Uchida

Softmax-Gradient Layer-wise Relevance Propagation (SGLRP) enhances the interpretability of Convolutional Neural Networks by generating class-discriminative pixel-wise attribution maps, effectively highlighting regions specific to a target object's classification. The method demonstrates improved object localization and reduced relevance attribution to non-target elements compared to prior LRP techniques.

Key insights

  • SGLRP's use of softmax gradients provides a probabilistic foundation for initial relevance, allowing for an adaptive and proportional penalty for non-target classes based on their predicted probabilities.
  • The method maintains the conservation property, ensuring the sum of relevances at the output layer is zero, balancing target and non-target contributions.
  • SGLRP is generally applicable and compatible with various deep CNN architectures, including VGG16, ResNet152, DenseNet121, and InceptionV3.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.