Reproduce “Guided Integrated Gradients: An Adaptive Path Method for Removing Noise” 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
Andrei Kapishnikov, Subhashini Venugopalan, Besim Avci, Ben Wedin, Michael Terry, Tolga Bolukbasi
Google Research introduces Guided Integrated Gradients, a method that addresses noisy attributions in Integrated Gradients by proposing an entirely new paradigm for path-based attribution methods. It constructs adaptive integration paths by iteratively moving features with the lowest absolute partial derivatives, resulting in cleaner and more focused attribution maps that consistently outperform standard Integrated Gradients on quantitative metrics.
Key insights
- Noisy attributions in Integrated Gradients stem from the accumulation of irrelevant gradient information along the standard straight-line integration path.
- A new framework, Adaptive Path Methods (APMs), allows the integration path to dynamically adjust based on the neural network function being explained.
- By iteratively moving only features with the lowest absolute partial derivatives, an adaptive path can avoid regions of high, potentially irrelevant gradients, leading to cleaner attributions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.