Software engineers require a large amount of data for building neural network-based software systems. The engineering of these data is often neglected, though, it is a critical and time-consuming activity. In this work, we present a novel software engineering approach for dataset augmentation using neural networks. We propose a rigorous process for generating synthetic data to improve the training of neural networks. Also, we demonstrate our approach to successfully improve the recognition of handwritten digits using conditional generative adversarial networks (cGAN). Finally, we shortly discuss selected important issues of our process, presenting related work and proposing some improvements.
Defining and managing teaching programs at universities or other institutions is a complex task for which there is not much support in terms of methods and tools. This task becomes even more critical when the time comes to obtain certifications w.r.t. official standards. In this paper, we present an on-going project called TESMA, whose objective is to provide an open-source tool dedicated to the specification and management (including certification) of teaching programs. An in-depth market analysis regarding related tools and conceptual frameworks of the project is presented. This tool has been engineered using a development method called Messir for its requirements elicitation and introduces a domain-specific language dedicated to the teaching domain. This paper presents the current status of this project and the future activities planned.
Neural network (NN) components are being increasingly incorporated into software systems. Neural network properties are determined by their architecture, as well as the training and testing datasets used. The engineering of datasets and neural networks is a challenging task that requires methods and tools to satisfy customers’ expectations. The lack of tools that support requirements specification languages makes it difficult for engineers to describe dataset and neural network recognition skill requirements. Existing approaches often rely on traditional ad hoc approaches, without precise requirement specifications for data selection criteria, to build these datasets. Moreover, these approaches do not focus on the requirements of the neural network’s expected recognition skills. We aim to overcome this issue by defining a domain-specific language that precisely specifies dataset requirements and expected recognition skills after training for an NN-based system. In this paper, we present a textual domain-specific language (DSL) called SEMKIS-DSL (Software Engineering Methodology for the Knowledge management of Intelligent Systems) that is designed to support software engineers in specifying the requirements and recognition skills of neural networks. This DSL is proposed in the context of our general SEMKIS development process for neural network engineering. We illustrate the DSL’s concepts using a running example that focuses on the recognition of handwritten digits. We show some requirements and recognition skills specifications and demonstrate how our DSL improves neural network recognition skills.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.