Treffer: Variation Control Support for ML-based Systems

After their successful debut in research, Machine Learning (ML) and Artificial Intelligence (AI) topics are now reaching the industry, where version control, software product lines, and especially DevOps and CI/CD are part of the day-to-day practice. The gap between these already established processes and the relatively new methods of ML opened up a new field that has been named ML DevOps in recent literature. However, this area is still missing common definitions and standards. This thesis wants to contribute to closing this gap by providing variation control support for ML projects. After analyzing, how variation control systems can contribute in the workflow of ML DevOps projects the main requirements are defined. When then extend the variation control system ECCO to support Python files and Jupyter Notebooks, two commonly used formats in machine learning projects. Subsequently, the implementation was tested for correctness and performance based on the defined use cases. The evaluation shows that the implemented adapter can successfully provide support for Jupyter Notebooks and Python files. Extensional and intensional correctness checks were conducted, extensional correctness could be demonstrated and manual intensional correctness showed the expected results and the measured quality of the intensional checkouts was within the acceptable range. The performance measurements were satisfying and identified areas with room for improvements in the implementation but also in the ECCO core implementation. While there are many other applications to tackle ML DevOps challenges, we took this approach to further extend the ECCO capabilities. During the thesis some of ECCOs drawbacks came to light, which we discussed in detail with suggestions for further ECCO improvements. The results and findings of this thesis open up new possibilities for ECCO and its support for ML projects and pave the way for further research and improvements. ; Nach ihrem erfolgreichen Debüt in der Forschung erreichen machinelles Lernen (ML) und ...