SESS: Saliency Enhancing with Scaling and Sliding

SESS improves the reliability of AI explanations across different scenarios: The method makes saliency maps (visual explanations of AI decisions) more robust to common real-world challenges like objects appearing at different sizes, multiple objects in one image, and distracting elements—issues that previously caused existing explanation methods to fail or produce unclear results.

This is a practical tool for responsible AI practitioners: SESS is designed as a "method and model agnostic extension," meaning it can be added to existing explanation techniques without requiring complete redesign, making it accessible for organizations already using AI systems who want to improve their explainability capabilities.

Better explanations support safer AI deployment: By generating "less noisy and more discriminative saliency maps," SESS helps stakeholders and auditors better understand what features an AI system is actually using to make decisions—critical information for identifying potential biases or errors before deployment in high-stakes applications.

Efficiency improvements make responsible AI more scalable: The paper introduces a pre-filtering step that removes uninformative maps, reducing computational costs while maintaining quality improvements—addressing a practical barrier to implementing explainability methods in resource-constrained settings.

Open science supports verification and trust: The authors released code publicly, enabling independent verification of results and allowing other researchers to build upon the work, which strengthens the credibility and adoption of the method in responsible AI practices.