Reproduce “Uncertainty Quantification for Gradient-based Explanations in Neural Networks” 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

Mihir Mulye, Matias Valdenegro-Toro

A pipeline to systematically quantify uncertainty in gradient-based explanations for neural networks was developed, integrating existing uncertainty estimation and explanation methods. This work introduced the Coefficient of Variation for assessing explanation confidence and adapted evaluation metrics, demonstrating that Guided Backpropagation typically generates more confident and less noisy explanations, particularly when combined with Flipout.

Key insights

  • It is feasible to systematically combine existing uncertainty estimation and explanation methods to quantify the uncertainty inherent in model explanations.
  • The Coefficient of Variation (CV) provides a concise and effective metric to represent explanation uncertainty, highlighting features where the attributed importance is less reliable.
  • Evaluation metrics can be adapted to assess the quality and robustness of explanation distributions, offering a more comprehensive measure than single-explanation evaluations.

Summary by alphaXiv.

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