SEMKIS-DSL: a Domain-Specific Language for Specifying Neural Networks’ Key-Properties

Neural networks are becoming increasingly part of today’s software systems. These neural networks are simplified models of the human brain that are mainly capable of learning from large datasets to compute some function based on recognized data. Engineering these datasets and these neural network-based software systems is a complicated and challenging task. Software engineers require methods and tools to engineer these datasets and neural networks for their customers and to satisfy their requirements. In general, they lack methods and tools to support the engineering of dataset and neural networks that satisfy the customer’s requirements. They follow traditional approaches consisting of time-consuming, imprecise and manual activities. Typically, these approaches are not supported by any tool that precisely analyse and specify the neural network’s recognition skills. In our previous work, we have introduced the notion of key-properties for describing the neural network’s recognition skills. In this paper, we define a domain-specific language to support our SEMKIS software engineering methodology for the dataset augmentation to improve network’s key-properties. We present the SEMKIS-DSL for the specification of the key-properties of a neural network. We illustrate the concepts of our DSL with a running example specifying a neural network for recognizing a digital meter counter state. This running example demonstrates a specification of the neural network’s key-properties using the SEMKIS-DSL and a successful improvement of the neural network’s recognition skills.