Integrating symbolic and statistical methods for testing intelligent systems: Applications to machine learning and computer vision
Integrating symbolic and statistical methods for testing intelligent systems: Applications to machine learning and computer vision
复制标题
集成符号和统计方法来测试智能系统：在机器学习和计算机视觉中的应用
DOI：
--
复制
发表时间：
2016
期刊：
影响因子：
--
通讯作者：
Sumit Kumar Jha
中科院分区：
文献类型：
--
作者：
A. Ramanathan;L. Pullum;Faraz Hussain;Dwaipayan Chakrabarty;Sumit Kumar Jha
嵌入式智能系统，从微小的可植入生物医学设备到大群的自主无人机系统，正在我们的日常生活中变得越来越普遍。虽然我们依赖于这些智能系统的完美功能，并且经常认为它们的行为正确性和安全性是理所当然的，但众所周知，很难生成测试用例，以暴露机器学习算法实现中的细微错误。因此，智能系统的验证通常通过研究它们在代表性数据集上的行为来实现，使用交叉验证和自举等方法。在本文中，我们提出了一种新的测试方法来研究智能系统的正确性。我们的方法使用符号决策过程，再加上统计假设检验，以验证机器学习算法。我们展示了我们如何利用我们的技术成功地识别微妙的错误（如位翻转）的k-means算法的实现。这些错误不容易通过标准验证方法（如随机测试）检测到。我们还使用我们的算法来分析使用OpenCV开源计算机视觉库构建的人体检测算法的鲁棒性。我们表明，即使当扰动是如此之小，相应的帧看起来与肉眼相同，人类检测实现可能无法检测到人类在扰动的视频帧。
Embedded intelligent systems ranging from tiny implantable biomedical devices to large swarms of autonomous unmanned aerial systems are becoming pervasive in our daily lives. While we depend on the flawless functioning of such intelligent systems, and often take their behavioral correctness and safety for granted, it is notoriously difficult to generate test cases that expose subtle errors in the implementations of machine learning algorithms. Hence, the validation of intelligent systems is usually achieved by studying their behavior on representative data sets, using methods such as cross-validation and bootstrapping. In this paper, we present a new testing methodology for studying the correctness of intelligent systems. Our approach uses symbolic decision procedures coupled with statistical hypothesis testing to validate machine learning algorithms. We show how we have employed our technique to successfully identify subtle bugs (such as bit flips) in implementations of the k-means algorithm. Such errors are not readily detected by standard validation methods such as randomized testing. We also use our algorithm to analyze the robustness of a human detection algorithm built using the OpenCV open-source computer vision library. We show that the human detection implementation can fail to detect humans in perturbed video frames even when the perturbations are so small that the corresponding frames look identical to the naked eye.