深度学习软件的质量问题是阻碍该技术广泛应用的一大瓶颈，已日渐成为业内共同关注的焦点。然而现阶段针对其测试驱动的调试技术研究仍极为薄弱，亟待深入开展。本项目旨在面向这一紧迫需求，研究形成一套完整系统的理论方法及工具。研究内容包括：(1)提出基于蜕变关系的软件输出验证充分性准则；(2)构建基于张量数据流的深度学习模型频谱；(3)设计基于多目标差分进化的深度学习delta调试算法；(4)开发相应的蜕变关系共享库与工具套件平台。本项目创新性地结合蜕变测试与张量数据流，构建与测试结果相耦合并具有良好解释性的白盒数据结构，以解决深度学习软件“测试谕言缺失”与“黑盒属性”两大特点所导致的调试困难，突破现有研究中的若干瓶颈。并通过将调试问题建模为多目标优化问题，从而形成完整的由测试驱动的自动化调试方案。本项目的研究符合当前紧迫需求，对于深度学习软件的质量保障具有重要的理论与应用价值。

Quality issue in deep learning software has been one of the significant bottlenecks in the promotion of relevant techniques, and has been a common concern in the community. However, current studies on automatic test-driven debugging for deep learning software are still very few, which urgently need to be studied. This project aims to meet the above requirement, studying and proposing a complete and systematic theory and tools, for automatic debugging on deep learning software. The research contents of this project include: (1) proposing metamorphic-relation-based testing criteria for output behavior coverage; (2) constructing tensor-flow-based spectrum for deep learning models; (3) designing delta debugging algorithm with multiple-object differential evolution; (4) implementing shared repository of metamorphic relations and building testing and debugging platform. This project innovatively combines metamorphic testing and tensor flow information, and builds well-descriptive white-box data structure that tightly couples testing results, in order to settle the difficulties from the “oracle problem” and “black-box property” of deep learning software, as well as to make a few breakthroughs. Besides, it will model the debugging task into a multiple-object optimization problem, and as a consequence form a complete automatic test-driven debugging solution. This project precisely meets current urgent requirement, and has very important theoretical value and practical value for deep learning software quality assurance.