Power Combination Network for Image Classification on Small Samples with Data Loss

• Published in: 2019 IEEE Intl Conf on Parallel & Distributed Processing with Applications, Big Data & Cloud Computing, Sustainable Computing & Communications, Social Computing & Networking (ISPA/BDCloud/SocialCom/SustainCom) pp. 1162 - 1167
• Main Authors: Song, Kexin, Yang, Fenglei, Sun, Zhuochen
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.12.2019
• Subjects: 6G mobile communication; Big Data; Cloud computing; data loss; Distributed processing; image classification; power combination; PowNet; small samples; Social computing
• DOI: 10.1109/ISPA-BDCloud-SustainCom-SocialCom48970.2019.00165

Neural networks with deeper convolutional layers are difficult to train. Improving information flows helps to simplify the training process and achieve the better accuracy. In this study, we propose a power combination network, which called PowNet. The PowNet uses fewer convolution layers than other networks before. We consider the flow of information in and between layers, rather than just maximizing the flow of information between layers. In each layer, we divide the feature maps into partitions, and use intra-layer power combination to learn the relationship between partitions to get the feature of each layer. We employ inter-layer power combination to learn the relationship between the features in different level and get the output. This comprehensive combination of structures allows the networks to learn more global information and optimize the propagation of gradients. Therefore, the networks using fewer convolutional layers can have better performance in unfavorable training or test environments. Furthermore, the networks can decrease the generalization errors. In the CIFAR-10 dataset, we show that networks using intra-layer and inter-layer power combinations can use fewer convolutional layers to achieve the similar classification accuracy. We separately evaluated the performance of PowNets and other networks in the original dataset, small samples training, and data loss, including training accuracy, test accuracy and error between them. It has shown that networks using power combination had more advanced performance. To verify the power combination's ability in reducing generalization errors, we also performed experiments in STL-10. The similar phenomenon was obtained.