EVENT DOSSIER
On September 7, 2026, a paper published on arXiv cs.CV pointed out that the field of model interpretation has long overlooked simple models as benchmarks. The study emphasized that when evaluating the interpretability of complex models, it is necessary to incorporate and compare these neglected baseline models to provide a more comprehensive and accurate analysis perspective.
本文提出现有模型解释方法普遍忽略基线，导致解释不精确甚至错误。作者重构了模型解释任务及原则，统一了基于梯度的方法、集成梯度（IG）方法和泰勒展开，明确了各方法的基线。研究分析了 IG、LayerCAM、ODAM 和 Difference Map 等方法的缺陷与误差，主张通过归因结果与目标之间的归因误差评估质量，摒弃基于边际效应或完美模型假设的 flawed 评估方式。作者修正了 IG 并开发了具有清晰合理基线的解释方法，支持任意层特征进行解释，不同层的解释差异反映了不同特征提取阶段的特征提取程度。