ReLESS: A Framework for Assessing Safety in Deep Learning Systems

Traditionally, software refactoring helps to improve a system's internal structure and enhance its non-functional features, such as reliability and run-time performance, while preserving external behavior including original program semantics. However, in the context of learning-enabled software systems (LESS), e.g., Machine Learning (ML) systems, it is unclear which portions of a software's semantics require preservation at the development phase. This is mainly because (a) the behavior of the LESS is not defined until run-time; and (b) the inherently iterative and non-deterministic nature of ML algorithms. Consequently, there is a knowledge gap in what refactoring truly means in the context of LESS as such systems have no guarantee of a predetermined correct answer. We thus conjecture that to construct robust and safe LESS, it is imperative to understand the flexibility of refactoring LESS compared to traditional software and to measure it. In this paper, we introduce a novel conceptual framework named ReLESS for evaluating refactorings for supervised learning by (i) exploring the transformation methodologies taken by state-of-the-art LESS refactorings that focus on singular metrics, (ii) reviewing informal notions of semantics preservation and the level at which they occur (source code vs. trained model), and (iii) empirically comparing and contrasting existing LESS refactorings in the context of image classification problems. This framework will set the foundation to not only formalize a standard definition of semantics preservation in LESS but also combine four metrics: accuracy, run-time performance, robustness, and interpretability as a multi-objective optimization function, instead of a single-objective function used in existing works, to assess LESS refactorings. In the future, our work could seek reliable LESS refactorings that generalize over diverse systems.