Educating AI Software Engineers: Challenges and Opportunities

To properly develop, test and use Artificial Intelligence (AI) applications, students and professionals need a well-defined AI software engineering (AISE) process and the appropriate tools. However, AISE, which is today mainly based on the use of deep learning (DL) neural networks, is still under development. This makes the education of AI software engineers particularly challenging, since there are no well-established methodologies, tools and practices, like in traditional Software Engineering (SE) education drawing on decades of experience and methods in all phases of software development, from requirements analysis over design and implementation to integration and testing. We analyze the main differences between traditional SE and AISE education and address challenges in AISE education. Our methodology is based on literature survey, analysis of own industry experience and statistical analysis of students works on AI applications. Our goal is to provide guidelines for an AISE process and propose a curriculum path for AISE education, which can be used to update a traditional SE curriculum. According to results of our analysis, the main challenges for the students are: Dealing with data and taking into account that algorithms change (learn) by data, selection and re-use of AI algorithms, model test, maintenance and automatizing the AISE process. We propose to address these challenges in SE curricula by teaching more statistical thinking with connections to software development, developing re-engineering capabilities, teaching a model-based AI approach and combining AI with virtual reality simulations. In the whole process, we consider an optimal division of work between humans and AI systems by explicitly including humans in the AISE loop.