首页 > 科技 > 计算机 > 计算机科学

深度学习与神经网络

作译者：赵金晶等

出版时间：2024-02 千字数：283 版次：01-01 页数：232

开本：16开装帧： I S B N ：9787121473739

换版：

纸质书定价：￥86.0

库存:无

分享到：

共有图书评论0条【查看评论摘要】

看了又看

{{data.bookName}}

￥{{data.price}}

内容简介

前言

上架建议

作者简介

编辑推荐

音视频专区

第1 章 绪论····················································································.1
1.1 人工智能·············································································.2
1.1.1 人工智能技术的发展历程···············································.3
1.1.2 人工智能技术的流派·····················································.9
1.2 深度学习与神经网络概述······················································.11
1.2.1 深度学习与神经网络技术的发展历程······························.11
1.2.2 深度学习与神经网络的前沿技术····································.16
1.3 深度学习系统架构·······························································.17
1.4 深度学习框架·····································································.19
1.5 深度学习的应用··································································.20
1.5.1 计算机视觉·······························································.20
1.5.2 语音语义··································································.21
1.5.3 自然语言处理····························································.22
1.6 人工智能潜在的安全风险······················································.22
1.6.1 数据层面的风险·························································.23
1.6.2 算法模型层面的风险···················································.23
1.6.3 智能计算框架层面的风险·············································.23
1.6.4 基础软硬件层面的风险················································.24
1.6.5 应用服务层面的风险···················································.24
本章小结··················································································.24
第2 章 预备知识············································································.25
2.1 相关数学基础·····································································.25
2.1.1 线性代数··································································.25
2.1.2 概率论·····································································.27
2.1.3 优化理论··································································.32
2.2 机器学习基础·····································································.34
2.2.1 机器学习算法的基本流程·············································.35
2.2.2 机器学习常用评价指标················································.36
2.2.3 典型机器学习算法······················································.41
2.3 实验环境基础·····································································.49
2.3.1 GPU 驱动的安装配置··················································.49
2.3.2 依赖环境的安装配置···················································.52
2.3.3 深度学习框架的安装配置·············································.54
2.3.4 集成开发环境的安装配置·············································.56
本章小结··················································································.57
第3 章 前馈神经网络······································································.58
3.1 感知器··············································································.58
3.1.1 单层感知器·······························································.58
3.1.2 多层感知器·······························································.61
3.1.3 前馈神经网络的基本结构·············································.62
3.2 激活函数···········································································.63
3.2.1 Sigmoid 函数·····························································.63
3.2.2 ReLU 函数································································.65
3.2.3 Tanh 函数·································································.67
3.2.4 Softmax 函数·····························································.68
3.3 误差反向传播·····································································.69
3.3.1 梯度下降法·······························································.69
3.3.2 链式法则··································································.70
3.3.3 反向传播··································································.72
本章小结··················································································.75
第4 章 深度模型的优化···································································.76
4.1 神经网络的优化问题····························································.76
4.1.1 局部最优和振荡陷阱···················································.76
4.1.2 梯度爆炸和梯度消失···················································.78
4.2 常见的优化算法··································································.80
4.2.1 梯度下降优化算法······················································.80
4.2.2 二阶优化算法····························································.81
4.3 自适应学习率算法·······························································.85
4.3.1 AdaGrad 算法····························································.85
4.3.2 RMSprop 算法···························································.86
4.3.3 Adam 算法································································.86
4.4 参数初始化方法··································································.87
4.4.1 随机初始化·······························································.87
4.4.2 Xavier 初始化····························································.88
4.4.3 He 初始化·································································.89
本章小结··················································································.90
第5 章 深度学习中的正则化·····························································.91
5.1 范数惩罚···········································································.91
5.1.1 L1 正则化··································································.91
5.1.2 L2 正则化··································································.93
5.2 数据集增强与噪声注入·························································.94
5.2.1 数据集增强·······························································.94
5.2.2 噪声注入··································································.95
5.3 提前停止···········································································.97
5.4 Dropout ·············································································.97
5.5 批归一化···········································································.99
本章小结·················································································.102
第6 章 卷积神经网络······································································103
6.1 卷积神经网络的发展历程·····················································.103
6.2 卷积神经网络的基本组成·····················································.104
6.2.1 卷积层····································································.105
6.2.2 池化层····································································.106
6.2.3 全连接层·································································.108
6.3 常见卷积神经网络结构························································.108
6.3.1 VGG 网络································································.109
6.3.2 GoogLeNet 网络························································.111
6.3.3 ResNet 网络·····························································.117
6.4 深度生成网络····································································.122
6.4.1 生成对抗网络···························································.122
6.4.2 深度卷积生成对抗网络···············································.123
6.5 图像分类案例····································································.134
6.5.1 步骤1：搭建环境······················································.135
6.5.2 步骤2：导入依赖库···················································.137
6.5.3 步骤3：获取数据······················································.137
6.5.4 步骤4：定义AlexNet 网络··········································.138
6.5.5 步骤5：模型初始化···················································.140
6.5.6 步骤6：模型训练······················································.140
6.6 目标检测案例····································································.143
6.6.1 步骤1：环境配置和模型下载·······································.144
6.6.2 步骤2：主函数解析···················································.151
6.6.3 步骤3：终端指令运行················································.152
本章小结·················································································.153
第7 章 循环神经网络·····································································.154
7.1 循环神经网络的基本原理·····················································.154
7.1.1 循环神经网络的原理··················································.154
7.1.2 双向循环神经网络·····················································.157
7.2 循环神经网络在实际中的应用···············································.158
7.2.1 文本生成·································································.159
7.2.2 语音识别·································································.159
7.2.3 机器翻译·································································.160
7.2.4 生成图像描述···························································.161
7.2.5 视频动作检测···························································.162
7.2.6 信号分类·································································.162
7.3 长短期记忆网络及其他门控循环神经网络································.163
7.3.1 长短期记忆网络························································.164
7.3.2 其他门控循环神经网络···············································.166
7.4 深度学习在文本和序列中的应用············································.167
7.4.1 文本数据处理···························································.167
7.4.2 文本分类和情感分析··················································.180
7.4.3 机器翻译·································································.180
7.4.4 命名实体识别···························································.182
7.5 卷积神经网络与循环神经网络···············································.183
7.5.1 卷积神经网络与循环神经网络的对比·····························.183
7.5.2 卷积神经网络与循环神经网络的组合应用·······················.184
7.6 案例：深度学习的诗歌生成··················································.185
7.6.1 步骤1：导入依赖库···················································.186
7.6.2 步骤2：读取数据······················································.187
7.6.3 步骤3：构造数据集···················································.188
7.6.4 步骤4：构造模型······················································.190
7.6.5 步骤5：训练过程······················································.192
7.6.6 步骤6：生成文本······················································.195
7.7 案例：基于LSTM 算法的股票预测·········································.196
7.7.1 步骤1：导入依赖库···················································.196
7.7.2 步骤2：获取并处理数据·············································.197
7.7.3 步骤3：构建预测数据序列··········································.199
7.7.4 步骤4：构建LSTM 网络············································.201
7.7.5 步骤5：训练网络······················································.202
7.7.6 步骤6：预测测试集···················································.203
7.8 案例：基于深度学习的文本分类············································.204
7.8.1 步骤1：项目入口······················································.205
7.8.2 步骤2：训练模块······················································.207
7.8.3 步骤3：验证和测试函数·············································.210
7.8.4 步骤4：数据预处理模块·············································.211
7.8.5 步骤5：定义模型······················································.215
7.8.6 步骤6：分类结果展示················································.218
本章小结·················································································.219

序
2006 年，杰弗里·辛顿和合作者发表了论文《一种深度置信网络的快速学
习算法》，标志着深度学习时代的到来，并由此掀起人工智能发展的第三次热
潮。尽管深度学习不能与人工智能画等号，但第三次热潮中诞生的令世人震惊的
绝大部分功能要归功于深度学习。可以预测，深度学习在未来很长一段时间内仍
是引领人工智能发展的核心技术。
深度学习的核心计算模型是深度人工神经网络。随着深度学习技术的成熟，
各类模型的规模越来越大，神经网络的层次越来越深，人工智能系统的“智能表
现”也越来越神奇。语音识别系统的发展突飞猛进、机器视觉识别超越人类、
人机聊天深度仿真等，各类应用竞相争艳。当然，深度学习的迅猛发展离不开
两大基石：按指数速度不断提升的强大计算能力和按指数速度不断积累的巨大
数据量。
第三次人工智能复兴之路与前两次的最大不同是，在深度学习技术赋能越来
越多的领域，不断产生颠覆性应用，并形成了成熟的商业模式，体现了真正的价
值。“旧时王谢堂前燕，飞入寻常百姓家”，今天人们谈到人工智能，几乎必谈深
度学习，曾经“阳春白雪”的机器学习，现在成为大家争相掌握的知识。
基于市场形式需求，赵金晶研究员等撰写了《深度学习与神经网络》一书。
该书介绍了深度学习和神经网络的基本原理和方法，用通俗易懂的语言和实例帮
助读者理解。内容聚焦各领域常用的模型和方法，如卷积神经网络、前馈神经网
络、模型优化问题等；基础理论部分尽可能缩减，只介绍神经网络基础数学知识
和机器学习基础理论，避免过多的概念对初学者造成困惑。书中实例的选取尽可
能与现实应用相结合，通过图像文本分类、股票预测、目标检测等场景的具体模
型，使读者直观地感受到深度学习方法解决现实问题的强大力量。
相信该书的出版能为进入该领域的广大读者提供有价值的参考。
赵金晶研究员曾是我的学生，看到昔日的学生著书立说我甚为高兴。特别是
获悉她和两位同事基于多年工作学习实践，用了近两年时间，多次推翻重写，再
经过五次大的修改最终成稿。我对这种认真治学的态度很是感动，特为之作序。
中国工程院院士
卢锡城
2023 年11 月22 日


前言
随着近几年数据量的积累、计算能力的增强及学习算法的优化，以神经网络
为代表的深度学习技术在人工智能的发展中独树一帜，在自然语言处理、目标识
别、机器翻译等领域都得到广泛应用。神经网络、图像识别、模式识别等相关课
程不仅是信息和计算机领域本科生的必修课，也成为医疗、设计、交通物流等其
他专业学生的课程。然而，若要深入理解并掌握深度学习与神经网络的理论和方
法，需要花费大量时间和精力。
神经网络的基本思想是通过对特征的逐层抽象以实现数据属性类别的表征。
在这一过程中，随着神经网络参数规模的扩大和复杂性的增加，在构建神经网络
模型的过程中会有许多问题和挑战，如过拟合和欠拟合、训练时空开销、调参、
可解释等。解决这些问题的好坏程度直接体现了模型的准确性和实用性。这些问
题涉及的知识面广，知识体系庞杂，如何将这些复杂的概念理论和方法用易于理
解的方式呈现给这个领域的初学者或工程实践者是笔者写这本书的初衷。
全书共分7 章。第1 章绪论，从介绍人工智能和深度学习技术的发展历程开
篇，梳理了人工智能不同技术流派的特点、深度学习的发展及前沿技术，在此基
础上介绍了本书后续章节所基于的深度学习系统架构和开发框架。第2 章介绍了
学习本书所应该掌握的相关预备知识，是掌握深度学习与神经网络方法所依赖的
数学理论基础，包括线性代数、概率论、优化理论及机器学习的基础知识。除此
之外，通过讲解案例详细指导读者如何构建深度学习实验环境，且后续学习过程
中能边学边实践。第3 章从前馈神经网络的基础模型——感知器出发，介绍了前
馈神经网络的基本结构及涉及的激活函数、梯度下降、误差反向传播等相关知
识。第4 章介绍了深度模型优化问题，讨论了神经网络优化中常见的病态问题，
如局部最优和振荡陷阱等。在此基础上介绍了各类深度学习中常见的优化算法的
基本原理、优缺点及适合的使用场景，最后介绍了构建深度学习模型中重要的参
数初始化方法。第5 章介绍了深度学习中的正则化，用于减少过拟合，并提高模
型的泛化能力，包括范数惩罚、数据集增强与噪声注入、提前停止、Dropout 和
批归一化。这些方法在深度学习中都得到了广泛应用，且相互结合可以取得更好
的效果。第6 章介绍了目前最流行的深度神经网络——卷积神经网络，对卷积神
经网络的发展历程、基本组成、常见的卷积神经网络结构，以及卷积神经网络在
计算机视觉领域的具体应用案例进行了详细介绍。第7 章对时序关联关系分析常
用的循环神经网络进行了介绍，讨论了循环神经网络衍生出的具有不同结构的各
类变体，还通过实际案例讨论了循环神经网络与卷积神经网络的结合应用模式。
虽然人们在深度学习和神经网络领域的研究已经取得了巨大的进步，但是仍
然有很多未知和挑战等待我们去探索。希望通过这本书，能够激发起读者对人工
智能技术的研究热情，继而在这个领域深入探索并取得新的突破。由于能力和
精力有限，书中内容难免出现差错，书中介绍的很多内容也是开放性的问题，
读者可通过电子邮箱zhjj0420@126.com 向笔者提出宝贵建议，一起探讨，共同
进步。
在此，我要感谢所有帮助我完成这本书的人，首先感谢我的导师卢锡城院
士，是他带领我进入了这个领域。他从高性能计算的独特角度来分析人工智能技
术的脉络，给了我不一样的视角，让我对人工智能和深度学习有了更深层次的理
解。其次，我要感谢电子工业出版社张正梅编辑对本书投入的巨大心力，感谢负
责编校和出版这本书的所有人员，是他们的专业知识和建议使这本书得以不断完
善。还要非常感谢我的学生钟文靖、王晨旭、张钊伟和张宏铮，他们是这本书的
第一读者，参与了本书部分章节的校审工作，提出了许多宝贵的意见和建议。感
谢复旦大学熊赟教授，国防科技大学李姗姗教授，军事科学院王蕾研究员、林白
研究员等专家对本书提出的宝贵意见。最后，深深感谢在本书撰写过程中给予我
理解、支持的家人，感谢我的父母、爱人和孩子一直以来对我的支持和鼓励，让
我在遇到各种打击和困难时，偶尔的自我否定和自我怀疑，在他们肯定和期翼的
目光中也会烟消云散，没有他们就不会有这本书的诞生。
道阻且长，行则将至。行而不辍，未来可期。
赵金晶
于北京，军事科学院
2023 年11 月6 日

查看更多 > 图书评论

暂无评论

发表图书评论

评论标题：
评论内容：
验证码：: 看不清楚
点击刷新
: 您还没有登录，请登录后再评论。
点击“登录”

购买过本书的顾客还买过

{{data.bookName}}

￥{{data.price}}