Inside the black box: A comprehensive survey of explainable AI benchmarks, protocols, technical analysis, and open gaps

초록

The rapid deployment of artificial intelligence (AI) in real-world pipelines has made explainable decision-making a core requirement for validation, auditing, and responsible use. This survey reviews explainable AI (XAI) methods, with attention to deployable systems, across seven domains: computer vision (CV), natural lan guage processing, tabular data, time series, graphs, reinforcement learning, and recommendation systems. To strengthen practical transparency for advanced deep models for visual intelligence, we synthesize evidence on visual interpretability, characterizing where models attend, how features are integrated across modalities, and how representations evolve across hierarchies. Guided by PRISMA-S search principles, we conducted a docu mented multi-library search across six source categories and applied explicit inclusion and exclusion criteria to assemble the literature analyzed here. The survey covers 189 works from 2014 to 2026, with primary synthesis drawn from the 2020 to 2026 period and foundational pre-2020 methods retained where they remain standard tools. We organize methods into seven families and, for CV, systematically catalog representative techniques by mechanism, model access, explanation output, and evaluation protocol. Across families, we discuss faithfulness, robustness, and computational cost, emphasizing the balance between efficiency and fidelity that governs prac tical deployment. To ground the visual interpretability analysis in a concrete system, we apply three established visualization techniques to our lightweight RGB-D saliency detection network (LiSalNet), producing Grad-CAM localization maps, multimodal fusion impact maps, and multiscale hierarchical visualizations with channel ac tivation statistics that reveal how cross-modal evidence and semantic abstraction emerge across network depth. Evidence from human studies is summarized to characterize reported effects on trust, reliance, task accuracy, and cognitive load.

키워드