Reproduce “iGOS++: Integrated Gradient Optimized Saliency by Bilateral Perturbations” 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
Saeed Khorram, Tyler Lawson, Fuxin Li
iGOS++ introduces an optimization-based saliency method that leverages bilateral perturbations and a novel regularization to generate coherent, high-resolution saliency maps. The approach demonstrates improved quantitative insertion scores and effectively uncovers model biases in a COVID-19 detection task, leading to enhanced classifier performance.
Key insights
- Considering both evidence removal and preservation through a bilateral perturbation strategy yields more robust and comprehensive saliency maps than deletion-only approaches, capturing a fuller set of influential features.
- Spatially adaptive regularization, such as Bilateral Total-Variance, is crucial for generating high-resolution saliency maps that are both accurate and human-interpretable, allowing sharp transitions at object boundaries while ensuring internal smoothness.
- High-fidelity saliency maps can serve as powerful debugging tools, effectively identifying subtle and unexpected model biases (e.g., overfitting to spurious text) that are otherwise undetectable, leading to more reliable AI systems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.