Facilitating Interdisciplinary Teams to Build Better AI-Based Systems

点击翻译按钮获取中文摘要

Intelligent software systems assist in every area of our lives, such as e-commerce sites, social media, and web searching. When building intelligent software systems with AI components, interdisciplinary teams consisting of (but not limited to) data scientists and software engineers need to work together. These groups have different goals and experience, often leading to friction in the process: Data scientists mainly work in the data exploratory phase to train a high-performing machine learning model, a heavily exploratory and iterative process. Then they deliver the resulting learning code and models to software engineers in order to integrate the model into production code in the production phase. Evidence shows that it is common that the code from the exploratory phase often needs to be refactored in order to accommodate production concerns, such as latency, scalability, and robustness. Additionally, once development and production models drift apart, it is non-trivial to incorporate feedback from the production phase for additional experimentation, often requiring significant coordination. This introduces mistakes, slows down the development process, and increases the need for coordination overhead. Currently, both big tech corporations and small start-up teams struggle with transitioning machine learning ideas into AI components that can be integrated into the software system seamlessly. The main goal of this research program is to foster collaboration and reduce friction between data scientists and software engineers, and provide support to the AI-based software development lifecycle. In particular, we aim to achieve three objectives: (1) Identifying context-specific collaboration pain points and best practices; (2) improving code quality and coding environment in the data exploration phase; and (3) facilitating collaboration between data scientists and software engineers. The main outcome includes the design and development of analysis infrastructure and interventions that support collaboration and system building. This research program will train 10 HQP (2 Ph.D., 3 MASc, 5 USR) through hands-on research practices in large-scale software analyses. The research activities involve in-depth user studies of software practitioners and AI experts, empirical investigation of the problem space, and rigorous design and evaluation of the methods to solve the problem. In addition, all HQP will gain skills and knowledge in software engineering, machine learning, and software development, and will have the chance to work on real, highly impactful software systems and build strong hands-on skills. Given the wide range of application scenarios, our research results can be applied to support collaboration and system building for teams focused on machine learning from different backgrounds, including established companies, start-ups, non-tech corporations, nonprofit, and research institutions in Ontario, in Canada, and internationally.