Saliency-Bench: A Comprehensive Benchmark for Evaluating Visual Explanations

Saliency maps lack standardized evaluation methods: The paper identifies a critical gap in responsible AI—while visual explanations (saliency maps) are widely used to show how AI models make decisions about images, there has been no consistent way to measure whether these explanations are actually accurate or trustworthy.

New benchmark tests explanations across real-world applications: Saliency-Bench evaluates visual explanations using eight diverse datasets covering practical scenarios including medical imaging (cancer diagnosis), scene recognition, and action classification, demonstrating that explanation quality varies significantly depending on the task and model architecture.

"Faithfulness" and "alignment" are measurable: The benchmark establishes standardized metrics to assess whether saliency maps truly reflect what a model is actually using to make decisions (faithfulness) and whether they match human understanding of important image features (alignment)—essential criteria for trustworthy AI systems.

Accessible tools democratize XAI evaluation: By providing an easy-to-use API and public benchmark, the research makes it practical for organizations to evaluate and compare different explanation methods without requiring specialized technical expertise, supporting broader adoption of responsible AI practices.

Explanation quality varies by method and architecture: Testing multiple saliency methods across different image classifier architectures reveals that no single explanation approach works best universally, highlighting the need for task-specific evaluation rather than one-size-fits-all solutions.