Towards Understanding the Challenges Faced by Machine Learning Software Developers and Enabling Automated Solutions
Modern software systems are increasingly including machine learning (ML) as an integral component. However, we do not yet understand the difficulties faced by software developers when learning about ML libraries and using them within their systems. To fill that gap firstly, this thesis reports on a detailed (manual) examination of 3,243 highly-rated Q&A posts related to ten ML libraries, namely Tensorflow, Keras, scikitlearn, Weka, Caffe, Theano, MLlib, Torch, Mahout, and H2O, on Stack Overflow, a popular online technical Q&A forum. Our findings reveal the urgent need for software engineering (SE) research in this area. The second part of the thesis particularly focuses on understanding the DNN bug characteristics. We study 2,716 high-quality posts from Stack Overflow and 500 bug fix commits from Github about five popular deep learning libraries Caffe, Keras, Tensorflow, Theano, and Torch to understand the types of bugs, their root causes and impacts, bug-prone stage of deep learning pipeline as well as whether there are some common antipatterns found in this buggy software.
While exploring the bug characteristics, our findings imply that repairing software that uses DNNs is one such unmistakable SE need where automated tools could be beneficial; however, we do not fully understand challenges to repairing and patterns that are utilized when manually repairing DNNs. So, the third part of this thesis presents a comprehensive study of bug fix patterns to address these questions. We have studied 415 repairs from Stack Overflow and 555 repairs from Github for five popular deep learning libraries Caffe, Keras, Tensorflow, Theano, and Torch to understand challenges in repairs and bug repair patterns. Our key findings reveal that DNN bug fix patterns are distinctive compared to traditional bug fix patterns and the most common bug fix patterns are fixing data dimension and neural network connectivity.
Finally, we propose an automatic technique to detect ML API misuses. We started with an empirical study to understand ML API misuses. Our study shows that ML API misuse is prevalent and distinct compared to non-ML API misuses. Inspired by these findings, we contributed Amimla (Api Misuse In Machine Learning Apis) an approach and a tool for ML API misuse detection. Amimla relies on several technical innovations. First, we proposed an abstract representation of ML pipelines to use in misuse detection. Second, we proposed an abstract representation of neural networks for deep learning related APIs. Third, we have developed a representation strategy for constraints on ML APIs. Finally, we have developed a misuse detection strategy for both single and multi-APIs. Our experimental evaluation shows that Amimla achieves a high average accuracy of ∼80% on two benchmarks of misuses from Stack Overflow and Github.
Committee: Hridesh Rajan (major professor), Gianfranco Ciardo, Gurpur Prabhu, Jin Tian, and Anuj Sharma