Enhancing Visual Feature Attribution via Weighted Integrated Gradients

Integrated Gradients has a significant baseline problem: The paper identifies that Integrated Gradients (IG), a popular method for explaining AI decisions in image recognition, produces unreliable explanations because it's overly sensitive to which reference image (baseline) is chosen—a limitation that affects the trustworthiness of AI explanations in real applications.

Not all reference images are equally useful: Current multi-baseline methods like Expected Gradients treat all reference images the same way, but this paper demonstrates that some baselines are more informative than others. Weighted Integrated Gradients (WG) automatically identifies and prioritizes the most useful reference images, improving explanation quality by up to 36%.

Better explanations come with a computational cost: While WG produces more reliable feature attributions across different types of AI models (both traditional convolutional networks and newer Transformers), it requires additional computation to evaluate baseline quality, making it a trade-off between explanation accuracy and processing speed rather than a faster solution.

The method maintains mathematical guarantees: WG preserves the core mathematical properties that make Integrated Gradients trustworthy, meaning the improvements in explanation quality don't come at the expense of losing important theoretical safeguards that ensure the method behaves predictably.

Improvements are consistent across different AI architectures: The gains from WG hold across multiple types of image recognition models and datasets, suggesting this approach could broadly improve how we explain AI decisions in computer vision applications.