本文为捷克布拉格查理大学（作者：Bc. Ján Vojt）的硕士论文，共104页。
深度神经网络是一种有效且通用的模型，能够完成各种各样的任务。本文主要研究了三种不同类型的深度神经网络——多层感知器、卷积神经网络和深度置信网络。所有讨论的网络模型都是在并行硬件上实现的，并且针对网络体系结构及其参数的各种选择进行了全面测试。作者对所实现系统的体系结构选择和优化都进行了详细的文档记录，最终的系统测试结果验证了该框架下的执行效率优势。本文的另一个重要部分还对支持深度神经网络的其它现有框架进行了额外的测试，对比测试表明，我们提出的设计框架优于多层感知器和卷积神经网络。
深度置信网络的性能稍微优于多达1000个隐藏神经元的RBM（Restricted Boltzmann Machine，受限玻尔兹曼机）层，但是对于更健壮的RBM层，与测试的竞争框架相比，其性能明显较低。
Deep neural networks represent an effectiveand universal model capable of solving a wide variety of tasks. This thesis isfocused on three different types of deep neural networks – the multilayerperceptron, the convolutional neural network, and the deep belief network. Allof the discussed network models are implemented on parallel hardware, andthoroughly tested for various choices of the network architecture and itsparameters. The implemented system is accompanied by a detailed documentationof the architectural decisions and proposed optimizations. The efficiency ofthe implemented framework is confirmed by the results of the performed tests. Asignificant part of this thesis represents also additional testing of otherexisting frameworks which support deep neural networks. This comparisonindicates superior performance to the tested rival frameworks of multilayerperceptrons and convolutional neural networks. The deep belief networkimplementation performs slightly better for RBM layers with up to 1000 hiddenneurons, but has a noticeably inferior performance for more robust RBM layerswhen compared to the tested rival framework.
1 引言
2 人工神经网络
3 具体实现
4 实验与测试
5 结论
下载英文原文地址：
http://page5.dfpan.com/fs/6l7c9jf212419219160/
更多精彩文章请关注微信号：