一句话：提出 ClusterXplain，提供 36 种基于聚类与迁移学习特征提取的流水线，用于识别 DNN 失败场景并辅助调试。
深度神经网络（DNN）在安全关键应用中一旦失败，工程师往往难以快速定位其失败场景与根因。本文提出 ClusterXplain，一个面向 DNN 组件调试的聚类工具。它延续先前工作 SAFE 的思路，利用迁移学习模型从诱发失败的图像中提取特征，并采用包括基于密度的方法在内的先进聚类算法对图像分组，从而识别 DNN 可能失败的不同场景。与 SAFE 仅关注单一流水线配置不同，ClusterXplain 系统性地探索特征提取方法、聚类算法与降维技术的多种组合，共提供 36 条调试流水线。实验结果表明，ClusterXplain 在场景识别上达到较高准确率，能为工程师诊断和缓解安全关键应用中的问题提供可操作的洞察，并附带演示视频。
We introduce ClusterXplain, a tool offering 36 distinct pipelines for debugging deep neural network (DNN) failures. Building on our previous approach, SAFE, ClusterXplain leverages transfer learning models to extract features from failure-inducing images. These features are used to cluster corresponding images using advanced algorithms, including density-based methods, to identify the distinct scenarios in which the DNN may fail. While SAFE focused on a single pipeline configuration, ClusterXplain explores diverse combinations of feature extraction methods, clustering algorithms, and dimensionality reduction techniques to improve DNN debugging. The results show that ClusterXplain achieves high accuracy in scenario identification, offering actionable insights to engineers to diagnose and mitigate issues in safety-critical applications. A demo video of ClusterXplain is available at https://youtu.be/DU94c-lIwys.