A psychophysics approach for quantitative comparison of interpretable computer vision models

Human-centered evaluation reveals gaps in automated metrics: The study found that quality metrics computed without human involvement "did not provide a consistent ranking of interpretability methods," suggesting that automated measures alone may not reliably indicate how useful AI explanations actually are to people.

Psychophysics methods offer a rigorous way to test interpretability: By using crowdsourced annotation tasks to measure both annotation accuracy and task completion time, the researchers demonstrated that techniques borrowed from psychology can provide "robust quality assessment of transparency in machine learning"—moving beyond purely qualitative comparisons.

Interpretability methods have measurable real-world impacts: Different explanation approaches produced measurable differences in how accurately humans completed tasks and how long those tasks took, indicating that the choice of interpretability method genuinely affects human performance, not just theoretical metrics.

We need better bridges between automated and human-centered evaluation: The disconnect between offline metrics and human-in-the-loop results highlights a critical gap in responsible AI: we cannot assume that interpretability methods that score well on automated benchmarks will actually help humans understand and work effectively with AI systems.

Scalability shouldn't come at the cost of validity: While the researchers acknowledge advantages of human-free evaluation (scalability, reproducibility, reduced bias), their findings suggest that responsible AI development requires periodic validation through human-centered studies to ensure automated metrics remain meaningful.