Analyzing source code with formal methods has a long history. However, most formal methods cannot capture “soft” aspects of source code, such as ambiguous information and hints within source code identifiers or comments.

In this talk, I will discuss a “black box” deep learning model and training method that finds and fixes seemingly simple but hard-to-find bugs. Specifically, they goal is to find bugs where there is a mismatch between the (latent) intent of the developer and source code. This necessitates models that reason over highly-structured data and code’s formal semantics. Here, structured deep learning models achieve state-of-the-art performance and the trained models find previously unknown bugs in open-source projects on GitHub.

I will conclude by discussing open challenges in this area.

I research deep learning methods for structured data, focusing on real-life applications and often to source code.