4. AI, ICT产业60年发展的总成果
AI: Overall outcome of 60 years of development in ICT
1956 1970s 1990s 2010s
AI popularity
第一次AI 冬天
AI Winter I
Moore’s Law 摩尔定律
第二次AI 冬天
AI Winter I I
AI 流行度
5. Railways
Iron steamship
Internal combustion
engine
Electricity
9000 BC~1000 AD 15th ~18th Century 19th Century 20th Century 21st Century
横跨整个经济的多种用途 巨大的技术互补性和溢出效应
Multiple uses across the economy Many technological complementarities and spillovers
https://en.wikipedia.org/wiki/General_purpose_technology
Richard G. Lipsey, etc., Economic Transformations: General Purpose Technologies and Long-Term Economic Growth
人工智能是一种新的通用目的技术(GPT)
AI is a new general purpose technology (GPT)
驯化植物
动物驯养
冶炼矿石
轮子
写作
青铜
铁
水轮
Domestication of plants
Domestication of animals
Smelting of ore
Wheel
Writing
Bronze
Iron
Water wheel
三帆帆船
印刷
工厂体系
蒸汽机
Three-masted sailing ship
Printing
Factory system
Steam engine
铁路
铁轮船
内燃机
电力
汽车
飞机
大规模生产
电脑
精益生产
互联网
生物技术
Automobile
Airplane
Mass production
Computer
Lean production
Internet
Biotechnology
商业虚拟化
纳米技术
Business virtualization
Nanotechnology
人工智能
(一组技术集合)
Artificial intelligence
(A set of technologies)
6. 改变将涉及所有行业
AI will change all industries
Public safety 公共
• Safe City
平安城市
• Intelligent transport
智慧交通
• Disaster prediction
灾害预警
Education 教育
• Personalization
个性内容
• Attention improvement
注意力提升
• Robo teacher
机器助教
Healthcare 健康
• Early prevention
早期预防
• Diagnosis assistance
协助诊断
• Precision cure
精准治疗
Media 媒体
• Real-time translation
实时翻译
• Abstraction
内容摘要
• Inspection
内容审核
Logistics 物流
• Routing planning
路径规划
• Monitoring
货物监视
• Auto sorting
自动分拣
Finance 金融
• Doc process
文档处理
• Real-time fraud prevention
实时防欺
• Up-sell
精准推荐
Pharmacy 制药
• Fast R&D
缩短周期
• Precise trial
精准试验
• Targeted medicine
精准药物
Insurance 保险
• Auto detection
高效鉴定
• Fraud prevention
欺诈预防
• Innovative service
产品创新
Retail 零售
• Staff-less shops
无人超市
• Real-time inventory
实时库存
• Precise recommendations
精准推荐
Manufacturing 制造
• Defect detection
品质检测
• Industrial internet
工业物联
• Predictive maintenance
预测维护
Telecom 电信
• Customer service
客户服务
• Auto O&M
网络维护
• Auto optimization
网络优化
Agriculture 农业
• Fertilization improvement
施肥优化
• Remote operation
远程作业
• Seeds development
高效育种
Oil & Gas 油气
• Localization
精准钻探
• Remote maintenance
远程维护
• Operation optimization
运营优化 For illustration only
所列各行业AI应用只是示例
9. Speech recognition: On par with
2017年发表的ML论文数machine learning papers in 201720k 20k,
# of AI papers keeping up with Moore’s law in past 8 years Moore’s Law,8年来AI论文数快速增长
Object detection: Outperforming humans 目标检测性能超越 水平
humans
人类
语音识别达到 水平人类
Translation: Approachinghumans 翻译水平逼近 水平人类
countries with national AI plans22+ 22+ 国家发布了AI计划
253+business and academic events in 2018 253+ 场AI商业和学术活动 (2018年)
new AI startups in 20171,100+ 1,100+ 新AI Startups诞生 (2017年)
AI-related M&As in 2017US$24 bn $24B AI 有关的M&A (2017年)
AI-related VC investments in 2017US$14 bn AI 相关的VC投资 (2017年)$14B
of B2B companies employ AI to augment sales processes
的企业已经投资或部署了AIof enterprises have invested in or deployed AI4%
of retailers have invested in and deployed AI 零售商已经投资或部署了AI
of higher education institutions use AI to augment experience 高等教育机构使用AI扩增学习体验
B2B 企业在销售流程中使用AI
of smart city implementations are using AI 部署的智慧城市中正在使用AI
of customer service operations integrated virtual assistants in 2017
咨询和系统集成服务项目是AI相关的 (2017年)of consulting and SI service projects were AI-related in 2017
客户支持服务业务操作中集成了VCA (2017年)
of smartphones with AI capabilities in 2017 的智能手机内置了AI (2017年)
of B2C/B2B2C apps in China include AI in 2018 中国市场的B2C/B2B2C应用内含了AI (2018年
– Available AI talent vs. Global demand1% 全球AI人才与需求之比
~2%
5%
10%
~5%
4%
2%
~10%
~10%
4%
1%
~2%
5%
10%
~5%
4%
2%
~10%
~10%
今天,令人兴奋的落差
Inspiring gaps we see today
13. Pervasive AI for all scenarios
Respects and protects user privacy
AI无处不在,任何场景
尊重和保护用户隐私
To BeAs Is
Mostly in cloud,
some at the edge
AI主要在云、少量在边缘
AI部署
AI
deployment
14. New algorithms that are data and
energy-efficient, secure, and explainable
数据高效(更少的数据需求)
能耗高效(更低的算力和能耗)
安全、可解释
To BeAs Is
Today’s basic algorithms
invented before the 1980s
主要算法
诞生于1980年代
算法
Algorithms
15. Automated / semi-automated
data labeling, data collection, feature
extraction, model design, training, etc.
自动化 / 半自动化
数据标注、数据获取、特征提取、模型设计和训练…
To BeAs Is
No labor,
no intelligence
没有“人工”就没有“智能”
AI自动化
AI
automation
16. Industrial-grade AI,
perform excellently in execution
工业级AI, “工作”优秀
To BeAs Is
Models perform
better in tests
模型性能“考试”优秀
面向实际应用
Practical
application
arXiv.org > cs > arXiv:1806.00451
Compute Science > Machine Learning
Do CIFAR-10 Classifiers Generalize to CIFAR-10?
Benjamin Recht, Rebecca Roelofs, Ludwig Schmid, Vaishaal Shankar
(Submitted on 1 Jun 2018)
“CIFAR-10分类器能否泛化到CIFAR-10”
18. Synergy between AI and cloud, IoT, edge
computing, blockchain, big data, databases, etc.
协同云、物联网、边缘计算、
区块链、大数据、数据库…
To BeAs Is
Inadequate integration with
other technologies
与其它技术集成不充分
多技术协同
Multi-tech
synergy
19. AI as a basic skill, supported
by one-stop platforms
由一站式平台支持的
基本技能
To BeAs Is
Only highly-skilled
experts can work with AI
一项需要高级技能的、
专家的工作
平台支持
Platform
support
20. Data scientists + Subject matter experts
+ Data science engineers
数据科学家、领域专家、数据科学工程师相互协作
To BeAs Is
Scarcity
of data scientists
数据科学家稀缺
人才获得
Talent
availability
21. To BeAs Is
AI: Mostly in cloud, some at the edge
AI主要在云、少量在边缘
10项改变, 开创未来
10 changes that will shape the future
Today’s basic algorithms invented before the 1980s
主要算法诞生于1980年代
No labor, no intelligence
没有“人工”就没有“智能”
Models perform better in tests
模型性能“考试”优秀
Updates not in real time
非实时更新
Inadequate integration with other technologies
与其它技术连接不充分
Only highly-skilled experts can work with AI
一项需要高级技能的、专家的工作
Scarcity of data scientists
数据科学家稀缺
Training in days or even months
训练需要数日、数月
Scarce & costly computing power
算力稀缺且昂贵
Pervasive AI for all scenarios. Respects and protects user privacy
AI无处不在,任何场景;尊重和保护用户隐私
Data and energy-efficient, secure, and explainable algorithms
数据高效、能源高效、安全、可解释的算法
Automated / semi-automated data labeling
自动化 / 半自动化数据标注
Industrial-grade AI, perform excellently in execution
工业级AI, “工作”优秀
Real-time, closed-loop system
实时闭环系统
Synergy between AI and cloud, IoT, edge computing, blockchain,
big data, databases, etc.
协同云、物联网、边缘计算、区块链、大数据、数据库…
AI as a basic skill, supported by one-stop platforms
由一站式平台支持的基本技能
Data scientists + Subject matter experts + Data science engineers
数据科学家、领域专家、数据科学工程师相互协作
Training in minutes or even seconds
训练只需几分钟、几秒钟
Abundant & affordable computing power
算力充裕且经济
22. 在计算视觉、自然语言处
理、决策推理等领域构筑数
据高效(更少的数据需
求) 、能耗高效(更低的算
力和能耗) ,安全可信、自
动自治的机器学习基础能力
Invest in
AI research
投资基础研究
Develop fundamental
capabilities for data &
power-efficient (i.e., less
data, computing, and
power needed), secure &
trusted, automated /
autonomous machine
learning in computer
vision, natural language
processing, decision /
inference, etc.
华为AI发展战略
Huawei’s AI strategy
打造面向云、边缘和端等全
场景的、独立的以及协同
的、全栈解决方案,提供充
裕的、经济的算力资源,简
单易用、高效率、全流程的
AI平台
Build a full-stack
AI portfolio
打造全栈方案
Deliver abundant and
affordable computing
power
Provide an efficient and
easy-to-use AI platform
with full-pipeline services
Adaptive to all scenarios,
both standalone and
cooperative scenarios
between cloud, edge, and
device
面向全球,持续与学术界、
产业界和行业伙伴广泛合作
Develop an open
ecosystem and talent
投资开放生态和人才培养
Collaborate widely with
global academia, industries,
and partners
把AI思维和技术引入现有产
品和服务,实现更大价值、
更强竞争力
Strengthen existing
portfolio
解决方案增强
Bring an AI mindset and
techniques into existing
products and solutions to
create greater value and
enhance competitive
strengths
应用AI优化内部管理,对准
海量作业场景,大幅度提升
内部运营效率和质量
Drive operational
efficiency
内部效率提升
Apply AI to massive
volumes of routine business
activities for better efficiency
and quality
23. 华为AI解决方案
Huawei’s AI portfolio
Full Stack
CANN
(Compute Architecture for Neural Networks)
Ascend
All Scenarios全场景
MindSpore TensorFlow PyTorch PaddlePaddle …
Application
Enablement
Framework
Chip Enablement
IP & ChipAscend-MaxAscend-MiniAscend-Tiny Ascend-LiteAscend-Nano
全栈
AI Applications AI 应用
Application enablement:
Full-pipeline services (ModelArts),
hierarchical APIs, and pre-integrated
solutions
CANN:
Chip operators library and highly
automated operators development
toolkit
Ascend:
AI IP and chip series based on
unified scalable architecture
MindSpore:
Unified training and inference
framework for device, edge, and
cloud (both standalone and
cooperative)
应用使能:
提供全流程服务(ModelArts), 分层
API和预集成方案
芯片算子库和高度自动化算子开发工具
基于统一、可扩展架构的系列化AI IP 和
芯片
支持端、边、云独立的和协同的统一训
练和推理框架
应用使能
框架
芯片使能
ModelArts
General APIs Advanced APIs Pre-integrated Solutions
HiAI Engine
HiAI Service
IP和芯片
消费终端
Consumer Device Public Cloud
公有云 私有云
Private Cloud
边缘计算 IoT 行业终端
Industrial IoT
Device
Edge
Computing
24. HiAI
CANN
(Compute Architecture for Neural Networks)
ModelArts
General APIs Advanced APIs Pre-integrated Solutions
Ascend
All Scenarios 全场景
MindSpore TensorFlow PyTorch PaddlePaddle …
Application
Enablement
Framework
Chip Enablement
IP & Chip
HiAI Engine
HiAI Service
Ascend-MaxAscend-MiniAscend-Tiny Ascend-LiteAscend-Nano
Full Stack
全栈
AI Applications AI 应用
应用使能
框架
芯片使能
消费终端
Consumer Device Public Cloud
公有云 私有云
Private Cloud
边缘计算 IoT 行业终端
Industrial IoT
Device
Edge Computing
IP和芯片
HiAI service 基于Cloud EI 部署
Full-stack portfolio for smart devices
HiAI services are deployed on Cloud EI
面向智能终端的全栈解决方案
ModelArts
General APIs Advanced APIs Pre-integrated Solutions
Ascend
MindSpore TensorFlow PyTorch PaddlePaddle …
HiAI Engine
HiAI Service
Ascend-MaxAscend-MiniAscend-Tiny Ascend-LiteAscend-Nano
消费终端
Consumer Device Public Cloud
公有云 私有云
Private Cloud
边缘计算
Edge Computing
公有云、私有云、混合云、边缘、IoT行业终端
Full-stack portfolio for organizations
(governments, enterprises, etc.)
Public / private / hybrid clouds, edge, and industrial IoT devices
面向组织(政府、企业等)的全栈解决方案
华为 AI 解决方案:HiAI 和 EI
Huawei’s AI portfolio: HiAI and EI
AI Applications AI 应用
Full Stack
全栈
Application
Enablement
Framework
Chip Enablement
IP & Chip
应用使能
框架
芯片使能
IP和芯片
CANN
(Compute Architecture for Neural Networks)
All Scenarios 全场景
EI
IoT 行业终端
Industrial IoT
Device
25. Half-Precision (FP16): 256 TeraFLOPS
Integer-Precision (INT8): 512 TeraOPS
128 Channel FHD Video Decoder – H.264/265
Ascend-Max
Architecture: Da Vinci
Ascend-Max
半精度 (FP16): 256 TeraFLOPS
整数精度 (INT8): 512 TeraOPS
128 通道 全高清 视频解码器 – H.264/265
架构: 达芬奇
Max Power: 350W
7nm
2019 Q2
7nm
2019 Q2
华为昇腾910
Google
TPU v2
Google
TPU v3
Ascend 910
45T
90T
125T
256T
Nvidia V100
* Normalized to 16-bit
FLOPS
1
2
3
4
华为昇腾910:单芯片计算密度最大
Ascend 910: Greatest computing density in a single chip
最大功耗: 350W
28. 华为昇腾,横跨全场景的最优TOPS/W
Ascend: Optimal TOPS/W across all scenarios
Earphone
-1mW
Ascend-Nano
Always-on
-10mW
Ascend-Tiny
Smartphone
1-2mW
Ascend-Lite
Laptop/PC/Module
3-10W
Ascend 310
Edge Server
10-100W
Multi-Ascend 310
Data Center
200+W
Ascend 910
Power
功率
T4
P4
TOPS/W
Ascend
*Normalized to 8-bit
29. AI Acceleration Module
AI 加速模块
Atlas 200
基于Ascend 310 AI产品
Ascend 310-based AI products
AI Acceleration Card
AI 加速卡
Atlas 300
AI Edge Station
AI 智能小站
Atlas 500
AI Appliance
AI 一体机
Atlas 800
Mobile Data Center
移动数据中心
MDC 600
31. 优化内部管理
SoftCOM AI
HiAI Cloud EI
FusionMind
Atlas/MDC
Internal
Management
Optimization
总结: 华为人工发展智能战略要点
Summary: Huawei’s AI strategy highlights
Research
+ Open Global Ecosystem
基础研究
+ 开放全球生态
+
Talent Development 人才培养
Full Stack
全栈
All Scenarios
全场景