Convergence of multiple deep neural networks for classification with fewer labeled data
- PDF / 1,494,084 Bytes
- 10 Pages / 595.276 x 790.866 pts Page_size
- 78 Downloads / 136 Views
ORIGINAL ARTICLE
Convergence of multiple deep neural networks for classification with fewer labeled data Chuho Yi 1 & Jungwon Cho 2 Received: 31 May 2020 / Accepted: 22 August 2020 # Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract With the advent of deep neural networks (DNNs) in the last two decades, tremendous developments have been made in many fields, such as image classification/recognition, voice recognition, and action recognition. These advanced DNNs require large amounts of labeled data, whose collection is costly and requires great effort. In this paper, we provide a convergence method for DNNs to solve some of these difficulties. First, we consider how to create labeled data using a generative adversarial network (GAN), one DNN method, and add additional networks to improve the quality of generated data. Then, we propose a convergence method for the DNNs and use a three-step evaluation to confirm this approach and show how to use the automatically generated data for training. With the method proposed in this paper, we hope that the manual work of labeling data can be reduced for many DNN applications. Keywords Deep neural networks (DNNs) . Generative adversarial network (GAN) . Convergence . Generation system of labeled data
1 Introduction Object classification is one of the most important technologies in the field of image vision. It is an essential element in almost all applications, including surveillance systems, autonomous driving, and smart factories [1]. Over the past 20 years, deep neural network (DNN)-based technology has improved dramatically, and it will continue to be studied in the future [2, 3]. Advanced DNN technology requires data gathering and data labeling processes. Data labeling work is often done manually, and the more complex the application is, the more data are needed [4] [5]. These manual operations require time and effort and can adversely affect performance if errors occur in the labeling operation.
* Jungwon Cho [email protected] Chuho Yi [email protected] 1
Department of Computer Information, Hanyang Women‘s University, Seoul 04763, South Korea
2
Department of Computer Education, Jeju National University, Jeju 63243, South Korea
In this paper, we use a generative adversarial network (GAN) that can learn existing data and create new, virtual data as a means of solving these difficulties. Goodfellow provided a generator that creates data similar to the original data using values generated by arbitrary random numbers based on existing data and a discriminator, which distinguishes these data from the original. He proposed a way to reduce the cost function while competing between the generator and the discriminator [6]. In this way, generated or “fake” data can be generated from any random number [7]. Yi proposed applying these methods to generation of training data and evaluated the performance of pedestrian classification in several stages [8]. However, the generated image objects used in the proposed method did not provide sufficient signal and shap
Data Loading...
