ExplainabilityTransparencyTrustworthinesscs.AI
Saliency strikes back: How filtering out high frequencies improves white-box explanations
Sabine Muzellec, Thomas Fel, Victor Boutin, Léo andéol, Rufin VanRullen, Thomas Serre
July 18, 2023
International Conference on Machine Learning
9 citations
1 influential
Key Takeaways
-
White-box explainability methods have a hidden flaw: Current attribution methods used to explain AI decisions rely on gradient signals that contain "high-frequency artifacts"—essentially noise that distorts the explanation of which inputs actually matter to the model's decision.
-
A simple filtering technique dramatically improves explanations: The researchers developed "FORGrad," which filters out these artifacts by removing high-frequency noise tailored to each model's specific characteristics, significantly improving the quality of explanations without requiring major computational changes.
-
Efficiency and accuracy don't have to be a trade-off: By fixing this limitation, FORGrad allows simpler, faster white-box methods to produce explanations as accurate as more complex "black-box" methods, meaning organizations can get better explanations without sacrificing computational efficiency.
-
This could accelerate responsible AI adoption: The researchers suggest their work will encourage wider use of white-box explainability methods by making them both faster and more reliable, which is important for building trustworthy AI systems that stakeholders can actually understand and verify.
Abstract
Attribution methods correspond to a class of explainability methods (XAI) that aim to assess how individual inputs contribute to a model's decision-making process. We have identified a significant limitation in one type of attribution methods, known as ``white-box" methods. Although highly efficient, as we will show, these methods rely on a gradient signal that is often contaminated by high-frequency artifacts. To overcome this limitation, we introduce a new approach called "FORGrad". This simple method effectively filters out these high-frequency artifacts using optimal cut-off frequencies tailored to the unique characteristics of each model architecture. Our findings show that FORGrad consistently enhances the performance of already existing white-box methods, enabling them to compete effectively with more accurate yet computationally demanding "black-box" methods. We anticipate that our research will foster broader adoption of simpler and more efficient white-box methods for explainability, offering a better balance between faithfulness and computational efficiency.