Specifying Key-properties to Improve the Recognition Skills of Neural Networks

Abstract

Software engineers are increasingly asked to build datasets for engineering neural network-based software systems. These datasets are used to train neural networks to recognise data. Traditionally, data scientists build datasets consisting of random collected or generated data. Their approaches are often costly, inefficient and time-consuming. Software engineers rely on these traditional approaches that do not support precise data selection criteria based on customer's requirements. We introduced a software engineering method for dataset augmentation to improve neural networks by satisfying the customer's requirements. In this paper, we introduce the notion of key-properties to describe the neural network's recognition skills. Key-properties are used all along the engineering process for developing the neural network in cooperation with the customer. We propose a rigorous process for augmenting datasets based on the analysis and specification of the keyproperties. We conducted an experimentation on a case study on the recognition of the state of a digital meter counter. We demonstrate an informal specification of the neural network's key-properties and a successful improvement of a neural network's recognition of the meter counter state.