Toward Robust and Adaptable Deep Learning Models of Code

点击翻译按钮获取中文摘要

The demand for software has never been so high and is expected to further increase in the years to come. Software systems have become a founding pillar of our heavily digitalized society, and have had a significant impact in numerous domains, including health care, banks, and autonomous vehicles. At the same time, the complexity of software systems tends to increase with time, especially as they must be prone to dealing with intense quantities of data as well as with other software and hardware components. This can have a massive impact on the software developers' efficiency. One solution to cope with these increasing demand and complexity is to use artificial intelligence, specifically deep learning, to automate some complex software engineering tasks. In the past years, deep-learning-based approaches have enabled us to learn potent models of the software world, and successfully use them, for example, to complete the code of a project under development or to search for pieces of code using a natural language query. Nonetheless, these models have limited applications outside the context in which they were trained. The main objective of this project is to make deep learning models more adaptive and robust to the evolution of data and application contexts. Instead of considering data as being stationary, our view is to consider data as streams continuously evolving through time.Many studies are projecting a significant growth in software developers employment from 2020 to 2030. There is currently a global race on developing innovative research-based tools to support software development using artificial intelligence. One of the main objectives is for each country to be able to deal with its shortage of software developers. For a country like Canada, projects like ours is essential to lead this race. Efficient and reliable software development is becoming a key component of fulfilling the short-term and long-term needs of society.