Learning to Find Bugs and Code Quality Problems - What Worked and What not?

The recent growth of open source repositories and deep learning models brought big promises for the next generation of programming tools that can automate or significantly improve the software development process. Yet, such tools are still rare and the machine learning components in them are not always apparent to their users. The current most useful techniques in machine learning for code are also not coming from the organizations such as Microsoft, Google, DeepMind, Facebook, OpenAI or nVidia that invested the most in deep neural techniques such as huge neural networks. This probably means that either many of these coding problems are significantly different from other hot topics in deep learning such as image processing or that it is much more difficult to collect datasets that would result in similarly successful tools. In this work, we study the results in the literature on the topic and discuss ways to address these shortcomings.