Studying an Architectural Pattern for Deep Learning training code - Collecting and Addressing Current Software Quality Issues within Academia and the Automotive Industry for Deep Learning

Abstract

Deep learning has become more popular throughout the years, consequently, an expansion of new developments within the field has occurred. As deep learning is mainly practiced by writing code, many established software engineering practices can be transferred to the field. While this has happened to some extent in some areas, like requirement engineering and MLOps, other subfields have lagged behind. Writing reusable and modular code is important for easy development, but there does not seem to exist a convention for how to write such code for deep learning training. Therefore, the architectural pattern MODLR was created and in this thesis, it was analyzed against found problems from practitioners of deep learning. One of the main goals of MODLR is to decouple loss code and to show the relevance of this focus, GitHub repositories were mined and automatically categorized projects based on their loss code, with the help of an LLM. The results show that MODLR is a good fit for an architectural pattern within the space of deep learning. As a bonus, it also shows one of the ways LLMs can be used to help research with automatic large-scale analysis of code.