SHF: Small: Deep Learning Software Repositories

College Of William And Mary, Williamsburg VA

Investigators

Abstract

Improvements in both computational power and the amount of memory in modern computer architectures, have enabled new approaches to canonical machine learning tasks. Specifically, these architectural advances have enabled machines, which are capable of learning deep compositional representations of massive data repositories. The rise of deep learning has ushered tremendous advances in several fields, and, given the complexity of software repositories, our hypothesis is that deep learning has the potential to usher new analytical frameworks and methodologies for Software Engineering research as well practice. The research program addresses three main goals by applying deep learning where conventional machine learning has been used before. First is the design of new models based on deep architectures for Software Engineering tasks. The project will develop deep software language models for sequence analysis tasks and deep information retrieval models for document analysis tasks. Second, the project will apply the internal representations to practical problems in Software Engineering by instantiating deep learning to support tasks such as code suggestion, improving software lexicons, model-based testing, code search and clone detection. Third, the project will conduct empirical evaluations designed to demonstrate ways of modeling software artifacts that will inform entirely novel suites of learned features that can be used from task to task. The move from traditional machine learning to deep learning will improve results in many software analysis tasks and in empirical Software Engineering research.