Timothy Elvira , Jaxon Selzer , Juan Ortiz Couder and Omar Ochoa , Embry-Riddle Aeronautical University, United States

Currently, the rise of software engineering, specifically requirements specifications, are being implemented in the design and workflow of ML applications, namely MLOps. Software Engineering has many untapped techniques, in both validation and verification, which could carry over into the ML context, providing a well engineered approach to developing ML models. One such technique is Equivalence Class Partitioning (ECP) which is a method of testing wherein testing a single test is valid for an ECP class, typically derived from a functional requirement. This paper explores a framework for deriving requirements, equivalence classes, and testing to develop a new form of equivalence class portioning, named ML-Behavior class partitioning to identify behavior of an ML model. This paper outlines the necessary changes of altering the traditional-SE ECP to accommodate the non-determinism and stochasticity of ML. To test the ML-Behavioral Class Partitioning testing (ML-BPC), an off-the-shelf YOLOv8 model is tested to examine behavior. Requirements are derived for the model using five image transforms: gaussian blurring, elastic distortion, translation, rotation, and brightness. The object is to identify the ML model’s behavior when incrementally increasing these image transforms, using ML-Behavioral Class Partitioning to identify the limits of the object detector by testing and fulfilling corresponding requirements.

Software Engineering, ML Engineering, Verification, Machine Learning, Object Detection.