摘要
The ensemble learning system based on neural network requires a large number of networks as the basic classifier, which makes the parameters and calculations of the system increase sharply. Integrating the neural network in depth can not only reduce the parameters and calculations of the network, but also improve the overall network performance. In this paper, a deeply integrated convolutional neural network (DICNN) was proposed, and several different integration methods were proposed for integrated learning of DICNN. The Mixup is used to train the DICNN because it uses multiple samples for training and it can be better combined with DICNN. A series of ablation experiments were done to prove that the training method of Mixup is equivalent to a regularization and data augmentation. Therefore, a different multi-sample training method as variations of the Mixup (Mixup-XL) can be used to train the DICNN.
延伸閱讀
- 傅品昇（2018）。A Study of M-Learning Using Convolutional Neural Network〔碩士論文，健行科技大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0022-0907201814245800
- 周揚賀（2015）。Using Deep Convolutional Neural Networks for Image Retrieval〔碩士論文，國立中央大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0031-0412201512063106
- 俞人暄（2016）。Design and Implementation of Convolutional Neural Network Processor〔碩士論文，國立交通大學〕。華藝線上圖書館。https://www.airitilibrary.com/Article/Detail?DocID=U0030-2212201712130539
- Kang, H. P. (2013). Improving Convolutional Neural Networks by Separable Filters [master's thesis, National Tsing Hua University]. Airiti Library. https://doi.org/10.6843/NTHU.2013.00814
- Weng, P. S. (2016). Architecture Design of Convolutional Neural Networks for Image Recognition [master's thesis, National Taiwan University]. Airiti Library. https://doi.org/10.6342/NTU201603732