A framework for Requirements specification of machine-learning systems

The rapid development of machine learning (ML) systems has raised many concerns over their quality. Due to the inherent complexity and uncertainty, most of the traditional quality assurance techniques have been challenged, including requirements specification. Current strategies mainly focus on model extraction from existing neural networks to improve interpretability and facilitate system analysis, but failing to include user expectations on the system. To handle the problem, this paper proposes a specification framework for ML requirements where each ML system is regarded as a set of snapshot systems along the evolvement process. There are 3 layers in the framework and the hierarchy indicates that higher-level models need to be built based on lower-level ones. The bottom layer consists of meta snapshot model and meta data model serving as the meta models for snapshot systems and data requirements respectively. The middle layer is for snapshot models each describing a snapshot system through relations between its outputs produced with different inputs. The top layer is a learning model capturing the evolvement process by transitions among snapshot models. These transitions are activated by data models instantiated from meta data model. We adopt the specification of a self-driving system to illustrate the framework.