本文为美国波士顿学院（作者：Serge Aleshin-Guendel）的论文，共31页。

近年来，机器学习与大型数据集相结合，出现了许多成功案例，导致人们对这一领域的兴趣增加。一类流行的机器学习模型是深度神经网络，在这种结构中，层叠的“神经元”被用于学习数据的近似表示。卷积神经网络（CNN）是一种特殊的模型，它已经成为大多数计算机视觉任务的标准。然而，关于如何最好地构建这些CNN几乎没有任何理论。本文的目的有两个。第一个目的是简要介绍有监督机器学习、神经网络和CNN领域。第二个目标是通过对网络的宽度、深度和感知场相关的结构特性的研究，探索如何最好地构建cnn。

Machine learning, in conjunction with largedata sets, has seen many success stories in recent years, leading to increasedinterest in the field. One popular class of machine learning models is deepneural networks, where stacked layers of “neurons” are used to learnapproximate representations of data. One particular model, the ConvolutionalNeural Network (CNN), is notable in that it’s become the standard in mostcomputer vision tasks. However, there is little to no theory surrounding how tobest build these CNNs. The aim of this thesis is two-fold. The first aim is toprovide a brief introduction to the field of supervised machine learning,neural networks, and CNNs. The second aim is to explore how to best build CNNs,through an examination of structural properties related to the width, depth,and receptive field of networks.

1. 引言

2. 神经网络

3. 实践中的神经网络训练

4. CNN结构研究

5. CNN结构研究：实验

6. 结论

更多精彩文章请关注公众号：