Analysis and Comparison of Pooling Layer in Deep Convolutional Neural Network

[[abstract]]人工智慧的概念是1955年美國科學家John McCarthy所提出，1958年時Rosenblatt提出了單層感知機的神經網路， 1984年日本學者Fukushima提出的神經認知機可以看作是卷積神經網路的前身，在1998年 LeCun, Bottou, Bengio, Haffner 提出了卷積神經網路的架構，也就是LeNet，而在2006年Hinton和Salakhutdinov 提出深度學習演算法-深度信念網路正式把類神經網路推上新的高峰，也讓本來已經被支援向量機替代的類神經網路重新被重視，也讓越來越多人開始重新研究如何讓類神經網路的精確度提高已達到更優秀的結果。類神經網路裡的深度卷積神經網路是目前很受歡迎的神經網路，因此本章決定探討它的架構並研究如何達到更好的識別成果。 深度卷積神經網路主要由輸入層、卷積層、池化層、激活函數和全連接層所組成。深度卷積神經網絡中的池化層主要是為了減少方差，降低計算複雜度，從鄰域中提取低級特徵並改善結果，減少過度擬合。本論文要以深度卷積神經網路裡LeNet-5的架構來為基底，並且使用Cifar-10資料庫來實驗，探討深度卷積神經網路裡池化層的不同對神經網路會造成的影響有哪些，並且找到能在減少影像維度的同時得到辨識時所需要的重要特徵。[[abstract]]The concept of artificial intelligence was proposed by American scientist John McCarthy in 1955. Since 1958, Rosenblatt proposed a neural network for perceptron. In 1984, the Neocognitron proposed by Japanese scientist Fukushima which could be regarded as the predecessor of CNN. Since 1998, LeCun, Bottou, Bengio, Haffner proposed the architecture of the Convolutional Neural Network (CNN), which is named LeNet, is the pioneer for CNN. In 2006, Hinton and Salakhutdinov proposed a deep learning algorithm - Deep Belief Network (DBN) has replaced the support vector machines (SVMs), and more and more people are beginning to re-examine how to improve the accuracy of neural networks. Deep convolutional neural networks are the most popular neural networks, so this paper decided to explore its architecture and study how to achieve better recognition results. Deep Convolutional Neural Network is mainly composed of an input layer, convolution layer, pooling layer, activation function and fully connected layer. The pooling layer in deep convolutional neural network is mainly to reduce the dimension to reduce over-fitting. This paper is based on LeNet-5 architecture and uses the Cifar-10 database to explore the different pooling in deep convolutional neural network that will make what impact.[[note]]碩