An integrated approach to software testing and learning

Software testing – the most commonly used approach for findings bugs – and machine learning – the most popular approach for extracting information from data – are two classic yet disparate approaches that serve largely different purposes and employ fairly unique techniques. Our thesis is that there is a novel bi-directional integration of the approaches that can make each of them more effective. In one direction, we define techniques based on the fundamentals of machine learning to enhance finding and removing bugs in various software systems. In the other direction, we define techniques based on the fundamentals of software testing to enhance the quality of machine learning systems. Our research consists of three thrusts. One, we define ePRFL, which uses the PageRank algorithm and Method-level aggregation technique to boost the accuracy of traditional techniques for localizing bugs in code. Two, we define DeepRoad, a GAN-Based Metamorphic Testing and Input Validation framework for DNN-based autonomous driving systems. Three, we define an analysis approach based on symbolic execution to study properties of deep neural networks. We embody the techniques in automated tools and evaluate them using standard benchmarks. The experimental results show the usefulness of our techniques.Electrical and Computer Engineerin