Contract-based Validation of Conceptual Design Bugs for Engineering Complex Machine Learning Software

Context. Modern software systems increasingly commonly contain one or multiple machine learning (ML) components. Current development practices are generally on a trial-and-error basis, posing a significant risk of introducing bugs. One type of bug is the "conceptual design bug," referring to a misunderstanding between the properties of input data and prerequisites imposed by ML algorithms (e.g., using unscaled data in a scale-sensitive algorithm). These bugs are challenging to test at design time, causing problems at runtime through crashes, noticeably poor model performance, or not at all, threatening the system's robustness and transparency. Objective. In this work, I propose the line of research I intend to pursue during my PhD, addressing conceptual design bugs in complex ML software from a prevention-oriented perspective. I intend to build open-source tooling for ML engineers that can be used to detect conceptual design bugs, enabling them to make quality assurances about their system design's robustness. Approach. We need to understand conceptual bugs beyond the status quo, identifying their types, prevalence, impacts, and structural elements in the code. We operationalize this knowledge into a tool that detects them at design time, allowing ML engineers to resolve them before running their code and wasting resources. We anticipate this tool will leverage contract-based validation applied to partial ML software models. Evaluation. We plan to evaluate the built tool two-fold using professional (industrial) ML software. First, we will study its effectiveness regarding bug detection at design time, identifying whether it fulfills its functional objective. Second, we will study its usability, identifying whether ML engineers benefit when tools like this are introduced into their ML engineering workflow.