SEMKIS: A CONTRIBUTION TO SOFTWARE ENGINEERING METHODOLOGIES FOR NEURAL NETWORK DEVELOPMENT

Today, there is a high demand for neural network-based software systems supporting humans during their daily activities. Neural networks are computer programs that simulate the behaviour of simplified human brains. These neural networks can be deployed on various devices e.g. cars, phones, medical devices...) in many domains (e.g. automotive industry, medicine...). To meet the high demand, software engineers require methods and tools to engineer these software systems for their customers.
Neural networks acquire their recognition skills e.g. recognising voice, image content...) from large datasets during a training process. Therefore, neural network engineering (NNE) shall not be only about designing and implementing neural network models, but also about dataset engineering (DSE). In the literature, there are no software engineering methodologies supporting DSE with precise dataset selection criteria for improving neural networks. Most traditional approaches focus only on improving the neural network’s architecture or follow crafted approaches based on augmenting datasets with randomly gathered data. Moreover, they do not consider a comparative evaluation of the neural network’s recognition skills and customer’s requirements for building appropriate datasets.
In this thesis, we introduce a software engineering methodology (called SEMKIS) supported by a tool for engineering datasets with precise data selection criteria to improve neural networks. Our method considers mainly the improvement of neural networks through augmenting datasets with synthetic data. SEMKIS has been designed as a rigorous iterative process for guiding software engineers during their neural network-based projects. The SEMKIS process is composed of many activities covering different development phases: requirements’ specification; dataset and neural network engineering; recognition skills specification; dataset augmentation with synthetized data. We introduce the notion of key-properties, used all along the process in cooperation with a customer, to describe the recognition skills. We define a domain-specific language (called SEMKIS-DSL) for the specification
of the requirements and recognition skills. The SEMKIS-DSL grammar has been designed to support a comparative evaluation of the customer’s requirements with the key-properties. We define a method for interpreting the specification and defining a dataset augmentation. Lastly, we apply the SEMKIS process to a complete case study on the recognition of a meter counter. Our experiment shows a successful application of our process in a concrete example.