Grant: Discovery Grants Program - Individual

Machine Learning (ML) algorithms have achieved tremendous advances in many application domains, such as image processing, medical diagnostics, and automated driving systems. To follow this technological hype, major cloud providers such as Amazon, Google and IBM have offered to developers a myriad of ML cloud services to ease the use of ML components in software systems. Using ML cloud services without the need to design or build the machine learning models from scratch has facilitated the usage and the integration of such components in software systems. The ever-increasing business requirements and the popularity of ML services have led developers of all skill levels to implement, maintain, and evolve ML-based systems rapidly. Hence, software engineers may not always follow good design and usage practices of ML cloud services. They may misuse such services which could significantly affect the quality of ML service-based systems and may consequently hinder their maintenance and evolution. The long-term objective of this research program is to build methods and tools for assessing and improving the software quality of ML service-based systems through the specification, detection and refactoring of ML service-specific misuses. We will first perform a domain analysis of good and bad practice usage of ML cloud services. We will then propose an automated approach to detect ML cloud service misuses in software systems. We will also study the impact of ML cloud service misuse on software quality. Finally, we will propose automatic refactoring solutions to correct ML cloud service misuses and improve the quality of ML service-based systems. This program is novel and original as we target software systems of prominent importance, i.e. ML service-based systems, which are recent, constantly evolving along their underlying frameworks, and are much more challenging to maintain than other traditional software systems. This work will provide software developers with techniques and tools to support their evolution tasks and to deliver a better experience to end users. Our methods and tools that will be developed in this program will enable both researchers and practitioners to improve the usage of ML services and raise awareness about potential misuses that can drastically impact the software quality. It will also save significant costs on the maintenance of such systems through automating the detection and refactoring of ML service misuses.