Two-level Formal Specifications for Deep Neural Networks
Yanzhao Xia, Shaoying Liu · International Journal of Software Engineering and Knowledge Engineering · 2025
Obtaining sufficient high-quality labeled data remains a critical challenge for training deep neural networks (DNNs). Recently, a specification-based method has been proposed to systematically define object characteristics for automated data generation. However, this approach typically relies on abstract descriptions, resulting in a gap between specifications and executable data generation. To address this issue, this paper proposes a two-level formal specification approach. Specifically, we apply the first-level specification to describe object characteristics and the second-level specification to define parameters and values suitable for data generation. This paper focuses on discussing both levels of specifications to facilitate human comprehension and machine handling to reduce the gap mentioned above. The performance of this approach is demonstrated through a case study on traffic sign recognition.