正版 创新工场讲AI课(从知识到实践) 创新工场DeeCamp组委会 电子

第1 章AI 赋能时代的创业······················································································1
1.1 中国AI 如何弯道超车····································································································2
1.2 AI 从“发明期”进入“应用期”··················································································9
1.2.1 深度学习助推AI 进入“应用期”···································································10
1.2.2 To B 创业迎来黄金发展期···············································································.11
1.2.3 “传统产业+AI”将创造巨大价值·····································································14
1.2.4 AI 赋能传统行业四部曲···················································································16
1.3 AI 赋能时代的创业特点·······························································································21
1.3.1 海外科技巨头成功因素解析·············································································21
1.3.2 科学家创业的优势和短板·················································································24
1.3.3 四因素降低AI 产品化、商业化门槛·······························································26
1.4 给未来AI 人才的建议··································································································30
第2 章AI 的产品化和工程化挑战·········································································35
2.1 从AI 科研到AI 商业化································································································36
2.2 产品经理视角―数据驱动的产品研发······································································40
2.2.1 数据驱动············································································································41
2.2.2 典型C 端产品的设计和管理············································································43
2.2.3 典型B 端产品解决方案的设计和管理·····························································46
2.2.4 AI 技术的产品化·······························································································48
2.3 架构设计师视角―典型AI 架构···············································································51
2.3.1 为什么要重视系统架构····················································································51
2.3.2 与AI 相关的典型系统架构··············································································53
2.4 写在本章最后的几句话································································································78
本章参考文献 ························································································································79
第3 章机器学习的发展现状及前沿进展 ······························································81
3.1 机器学习的发展现状····································································································82
3.2 机器学习的前沿进展····································································································85
3.2.1 复杂模型············································································································85
3.2.2 表示学习············································································································90
3.2.3 自动机器学习····································································································95
第4 章自然语言理解概述及主流任务 ··································································99
4.1 自然语言理解概述······································································································100
4.2 NLP 主流任务·············································································································100
4.2.1 中文分词··········································································································101
4.2.2 指代消解··········································································································102
4.2.3 文本分类··········································································································103
4.2.4 关键词（短语）的抽取与生成·······································································105
4.2.5 文本摘要··········································································································107
4.2.6 情感分析··········································································································108
本章参考文献·····················································································································.111
第 5 章机器学习在 NLP 领域的应用及产业实践···············································115
5.1 自然语言句法分析·····································································································.116
5.1.1 自然语言句法分析的含义与背景··································································.116
5.1.2 研究句法分析的几个要素··············································································.117
5.1.3 句法分析模型举例··························································································121
5.2 深度学习在句法分析模型参数估计中的应用····························································125
5.2.1 符号嵌入··········································································································126
5.2.2 上下文符号嵌入······························································································129
本章参考文献······················································································································131
第 6 章计算机视觉前沿进展及实践 ····································································133
6.1 计算机视觉概念··········································································································134
6.2 计算机视觉认知过程··································································································136
6.2.1 从低层次到高层次的理解···············································································137
6.2.2 基本任务及主流任务······················································································138
6.3 计算机视觉技术的前沿进展·······················································································141
6.3.1 图像分类任务··································································································141
6.3.2 目标检测任务··································································································148
6.3.3 图像分割任务··································································································151
6.3.4 主流任务的前沿进展······················································································155
6.4 基于机器学习的计算机视觉实践···············································································164
6.4.1 目标检测比赛··································································································164
6.4.2 蛋筒质检··········································································································167
6.4.3 智能货柜··········································································································170
本章参考文献······················································································································173
第 7 章深度学习模型压缩与加速的技术发展与应用·········································175
7.1 深度学习的应用领域及面临的挑战···········································································176
7.1.1 深度学习的应用领域······················································································176
7.1.2 深度学习面临的挑战······················································································178
7.2 深度学习模型的压缩和加速方法···············································································180
7.2.1 主流压缩和加速方法概述···············································································180
7.2.2 权重剪枝··········································································································182
7.2.3 权重量化··········································································································192
7.2.4 知识蒸馏··········································································································199
7.2.5 权重量化与权重剪枝结合并泛化···································································200
7.3 模型压缩与加速的应用场景·······················································································201
7.3.1 驾驶员安全检测系统······················································································202
7.3.2 高级驾驶辅助系统··························································································202
7.3.3 车路协同系统··································································································203
本章参考文献······················································································································204
第 8 章终端深度学习基础、挑战和工程实践·····················································207
8.1 终端深度学习的技术成就及面临的核心问题····························································208
8.1.1 终端深度学习的技术成就···············································································208
8.1.2 终端深度学习面临的核心问题·······································································209
8.2 在冗余条件下减少资源需求的方法··········································································.211
8.3 在非冗余条件下减少资源需求的方法·······································································213
8.3.1 特殊化模型······································································································214
8.3.2 动态模型··········································································································215
8.4 深度学习系统的设计··································································································216
8.4.1 实际应用场景中的挑战··················································································216
8.4.2 实际应用场景中的问题解决···········································································217
8.4.3 案例分析··········································································································219
本章参考文献······················································································································224
第 9 章DeeCamp 训练营最佳商业项目实战·······················································225
9.1 方仔照相馆―AI 辅助单张图像生成积木方头仔···················································227
9.1.1 让“AI 方头仔”触手可及·············································································227
9.1.2 理论支撑：BiSeNet 和Mask R-CNN ·····························································229
9.1.3 任务分解：从图像分析到积木生成的实现····················································231
9.1.4 团队协作与时间安排······················································································237
9.2 AI 科幻世界―基于预训练语言模型的科幻小说生成系统····································242
9.2.1 打造人机协作的科幻小说作家·······································································242
9.2.2 理论支撑：语言模型、Transformer 模型和GPT2 预训练模型·····················243
9.2.3 从“找小说”到“写小说”的实现步骤························································247
9.2.4 团队协作与时间安排······················································································250
9.3 宠物健康识别―基于图像表征学习的宠物肥胖度在线检测系统·························254
9.3.1 人人都能做“养宠达人”···············································································254
9.3.2 理论支撑：表征学习、人脸识别原理和ArcFace 损失函数·························257
9.3.3 任务分解：从数据收集到肥胖度检测···························································259
9.3.4 团队协作与时间安排······················································································262
9.4 商品文案生成―基于检索和生成的智能文案系统················································265
9.4.1 智能内容生成··································································································265
9.4.2 理论支撑：Word2Vec 词嵌入、预训练语言模型BERT 和Seq2Seq
文本生成··········································································································266
9.4.3 任务分解：“寻章摘句”和“文不加点”······················································269
9.4.4 团队协作与时间安排······················································································273
本章参考文献······················································································································276

主要作者（按章节顺序排序） 李开复 李开复博士于2009 年创立创新工场，担任董事长兼首席执行官，专注于科技创新型的投资理念与最前沿的技术趋势。十多年来创新工场已经投资逾400个创业项目，管理总额约160 亿人民币的双币基金。2016 年秋季创办创新工场人工智能工程院，致力于利用最前沿的AI 技术为企业提供人工智能产品与解决方案。 在此之前，李开复博士曾是谷歌中国全球副总裁兼大中华区总裁，担任微软全球副总裁期间开创了微软亚洲研究院，并曾服务于苹果、SGI 等知名科技企业。 李开复在美国哥伦比亚大学取得计算机科学学士学位，以zui gao荣誉毕业于卡耐基梅隆大学获得博士学位。同时，李开复获得香港城市大学、卡耐基梅隆大学荣誉博士学位。李开复获选为美国电机电子工程师学会（IEEE）的院士，并被《时代》杂志评选为2013影响全球100位年度人物之一，《Wired 连线》本世纪推动科技全球25位标杆人物，2018亚洲商界领袖奖等殊荣，并出任世界经济论坛第四次工业革命中心的AI 委员会联席主席。李开复博士发明过十项美国专利，发表逾百篇专业期刊或会议论文，并出版过十本中文畅销书。 王咏刚 王咏刚，现任创新工场CTO人工智能工程院执行院长，加入创新工场前担任谷歌主任工程师和高级技术经理超过十年，在谷歌参与或负责研发的项目包括桌面搜索、谷歌拼音输入法、产品搜索、知识图谱、谷歌首页涂鸦（Doodles）等，在知识图谱、分布式系统、自然语言处理、HTML5动画和游戏引擎等领域拥有丰富的工程研发经验。目前专注于人工智能前沿科技的工程化与商业化，以及人工智能高端人才的培养，作为联合创始人，创立了人工智能商业化公司创新奇智，同时也是人工智能高端应用型人才培养项目DeeCamp的发起者。 张潼 张潼博士，机器学习领域国际著名专家，拥有美国康奈尔大学数学和计算机双学士学位，以及斯坦福大学计算机硕士和博士学位，在香港科技大学数学系和计算机系任教。曾经担任美国新泽西州立大学终身教授，IBM研究院研究员和雅虎研究院主任科学家，百度研究院副院长和大数据实验室负责人，腾讯AI Lab主任。他曾参加美国国家科学院大数据专家委员会，负责过多个美国国家科学基金资助的大数据研究项目，此外还是美国统计学会和国际数理统计学会fellow，并担任NIPS、ICML、COLT等国际顶级机器学习会议主席或领域主席，以及PAMI, JMLR, 和Machine Learning Journal等国际一流人工智能期刊编委。 宋彦 宋彦博士，香港中文大学（深圳） 数据科学学院副教授，创新工场大湾区研究院首席科学家。历任微软、腾讯研究员及首席研究员，创新工场大湾区研究院执行院长等职，是 “微软小冰”项目的创始团队成员之一，其研究方向包括自然语言处理、信息检索和抽取、文本表征学习等。 屠可伟 屠可伟博士，上海科技大学信息科学与技术学院长聘副教授、研究员、博士生导师。研究方向包括自然语言处理、机器学习、知识表示、计算机视觉等人工智能领域，侧重于研究语言结构的表示、学习与应用。 张发恩 张发恩，创新奇智联合创始人，创新奇智公司CTO， 宁波诺丁汉大学客座教授。2008年毕业于中国科学院软件研究所，同年加入微软，负责Office相关软件产品的研发工作；2010年，入职Google，作为核心研发人员，主导和参与Google搜索引擎、Google知识图谱等相关工作；2015年，加入百度，作为百度云早期创始团队成员之一，曾任百度云计算事业部技术委员会主席，百度云计算事业部大数据和人工智能主任架构师。他在IT行业拥有十几年技术研发和管理经验，涉及企业级软件、室内地图定位与导航、互联网搜索引擎、全领域知识图谱、大数据计算与存储、机器学习、深度学习、机器视觉等众多领域。工作期间获得10余项美国专利，70余项中国专利，发表过多篇顶级会议学术论文。 唐剑 唐剑博士，滴滴智能控制首席科学家，AI Labs 负责人兼算法委员会主席，IEEE Fellow和ACM杰出科学家，领导滴滴在智能物联网、计算机视觉和自动驾驶方向上的研发。 他在国际顶级期刊和会议上发表了160多篇学术论文，拥有多项发明专利，在边缘智能、AI驱动的系统控制和群智感知方向上做出开创性贡献，并多次获得最佳论文奖, 其中包括通信网络领域的zui gao论文奖2019 年度IEEE 通信学会William R. Bennett Prize和IEEE车载技术学会2016年度最佳车载电子论文奖。 目前还担任中国电子学会物联网专委会专家委员、新一代人工智能产业技术创新战略联盟专家委员会委员、IEEE车载技术学会杰出演讲人、以及IEEE通信学会交换和路由技术委员会主席。 张弥 张弥博士，现任密歇根州立大学副教授，2006年毕业于北京大学，2013年获得美国南加州大学博士学位，2013-2014年在美国康奈尔大学任博士后。主要研究领域包括终端深度学习、自动机器学习、联邦学习和机器学习系统。本人及其研究团队在2019年Google MicroNet Challenge 全球竞赛中获得CIFAR-100赛道第四名 (北美第一名)，在 2017年NSF Hearables Challenge全球竞赛中获得第三名，在2016年NIH Pill Image Recognition Challenge全球竞赛中获得冠军。 吴佳洪 吴佳洪，创新奇智高级研究员，2017年毕业于北京大学。曾带队获得Pascal Voc 世界冠军、Cityscapes 实例分割冠军、ADE20K物体分割冠军等；曾负责创新奇智无人货柜项目的算法工作，专注于计算机视觉领域，发表过多篇CVPR论文。 刘宁 刘宁博士，滴滴资深研究员，毕业于美国东北大学计算机工程系。研究领域包括深度增强学习、深度模型压缩与加速、边缘计算等。在国际顶级期刊和会议AAAI, MICRO, ASPLOS，ISCA等发表学术论文20余篇，发表多项发明专利。

AI领域世界级科学家张亚勤、周明、周志华力荐；AI前沿+可落地的产业实践AI领域世界级科学家张亚勤、周明、周志华力荐；AI前沿+可落地的产业实践1.李开复、王咏刚、张潼等学术界、产业界领军人物全新梳理AI知识体系；2.AI领域国际科学家周明、张亚勤、周志华力荐；3.创新工场新创“知识授课+产业实践”学习模式；4.完整呈现从学科知识到工程实践的课程设计，实践项目来自产业界真实实践课题和训练数据；5.本书旨在将AI学习者、从业者从一招一式的“点滴型学习”转变为能够博取百家、融会贯通的“系统型学习”。

人工智能是第四次工业革命的技术基石，许多产业将被它改变，它也将带来许多新的产业—自动驾驶、工业互联网、AI医疗……我希望年轻的工程师们以饱满的精神迎接这样一个充满挑战的未来，我相信这本书能帮你做好准备。
—张亚勤 清华大学智能科学讲席教授、智能产业研究院院长什么是学习AI的最佳途径呢？我认为，如果有一个能够让学生体会到“耳到、眼到、口到、手到、心到”的学习环境无疑是最好的，而仅凭大学课堂很难做到这一点。
创新工场的DeeCamp训练营恰恰提供了这样一个环境—在短短几周时间里，学生通过大师讲座(耳到)、参观黑科技(眼到)、互相交流(口到)和编程实现(手到)，体会到AI系统设计理念和创新思维(心到)，从而激发了对AI的兴趣，走进了AI的殿堂。每一位学生都加入了一个小组，与队友合作，针对引人入胜的实际场景，把自己所学的知识融会贯通，把自己和队友的能力串联起来，看到项目一天天长大，从最初的一个好奇的想法落地成一个真正动起来的系统，每一刻都有豁然开朗之体味。这真是一个奇幻之旅。
—周明 创新工场首席科学家、微软亚洲研究院前副院长创新工场的DeeCamp训练营邀请了名师授课，在人工智能关键技术、工程实践、产品研发、创业转化方面都有涉及，对创新创业人工智能人才培养颇有裨益。诸多培训内容汇入本书，值得感兴趣的读者一读。
—周志华 南京大学教授、欧洲科学院外籍院士

创新工场于 2017 年发起了面向高校在校生的DeeCamp 人工智能训练营(简称DeeCamp训练营)，训练营内容涵盖学术界与产业界领军人物带来的全新AI 知识体系和来自产业界的真实实践课题，旨在提升高校AI 人才在行业应用中的实践能力，以及推进产学研深度结合。 本书以近两年 DeeCamp 训练营培训内容为基础，精选部分导师的授课课程及有代表性的学员参赛项目，以文字形式再现训练营"知识课程+产业实战”的教学模式和内容。全书共分为9 章，第1 章、第2 章分别介绍AI 赋能时代的创业、AI 的产品化和工程化挑战；第3 章至第8 章聚焦于AI 理论与产业实践的结合，内容涵盖机器学习、自然语言处理、计算机视觉、深度学习模型的压缩与加速等；第9 章介绍了 4 个优秀实践课题，涉及自然语言处理和计算机视觉两个方向。

适读人群 ：本书适合 AI 相关专业的高校在校生及AI 行业的工程师使用，可作为他们了解AI 产业和开拓视野的读物。

1.李开复、王咏刚、张潼等学术界、产业界领军人物全新梳理AI知识体系；

2.AI领域国际科学家周明、张亚勤、周志华力荐；

3.创新工场新创“知识授课+产业实践”学习模式；

4.完整呈现从学科知识到工程实践的课程设计，实践项目来自产业界真实实践课题和训练数据；

5.本书旨在将AI学习者、从业者从一招一式的“点滴型学习”转变为能够博取百家、融会贯通的“系统型学习”。