PowerPoint 演示文稿

Size: px

Start display at page:

Download "PowerPoint 演示文稿"

代邬
5 years ago
Views:

1 亚马逊人工智能与深度学习张侠 AWS 首席云计算企业顾问

2 日程人工智能与机器学习亚马逊人工智能 AI MXNet Polly Rekonization Lex

3 奇点来临奇点本是物理学术语, 原指时空中的一个普通物理规则不适用的临界点美国未来学家雷蒙德库兹韦尔用奇点来指电脑智能超越人脑智的那个神妙时刻我认为任何一种对心灵的冲击都比不过一个发明家亲眼见证人造大脑变为现实 -- 尼古拉特斯拉 1896

4 机器智能机器智能 ( 人工智能 ) 是研究开发用于模拟延伸和扩展人的智能的理论方法技术及应用系统的一门新的技术科学机器学习专家系统智能搜索智能控制定理证明游戏博弈机器人学机器视觉自动程序设计图像识别理解语音识别理解遗传编程

5 人工智能 : 机器学习机器学习专门研究计算机怎样模拟或实现人类的学习行为, 以获取新的知识或技能年代的理论和算法数量级增加的海量数据更强大更便宜的计算能力基于深度神经网的算法发明分布式的随机梯度下降法 SGD 等训练方法更多的大数据更强的硬件更好的算法

6 神经网络 : 多层感知输出, y 输出层 y = sigmoid(a o h 2 +b o ) 隐层 h 2 = sigmoid(a 2 h 1 +b 2 ) 隐层 h 1 = sigmoid(a 1 x+b 1 ) 输入层 x 输入, x

7 深度学习深度学习是基于于人工神经网络, 机器学习的一个分支含多隐层多层感知器的学习结构数据通过多层的非线性的转换目标 : 获取最有效的传递函数的转变参数方法 : 有监督学习与无监督学习之分等

8 深度学习许多领域的诸多应用都通过深度学习而得到重要改善图像理解语音识别自然语音处理自动化深度学习过去十年快速成长

9 亚马逊的人工智能应用

11 亚马逊人工智能全公司上千员工致力于人工智能发现与搜索物流配送提高现有产品定义新的产品类型将机器学习引入诸多领域

12 用户在 AWS 上部署 AI 示例 Zillow Zestimate (using Apache Spark) Howard Hughes Corp Lead scoring for luxury real estate purchase predictions FINRA Anomaly detection, sequence matching, regression analysis, network/tribe analysis Netflix Recommendation engine Pinterest Image recognition search Fraud.net Detect online payment fraud DataXu Leverage automated & unattended ML at large scale (Amazon EMR + Spark) Mapillary Computer vision for crowd sourced maps Hudl Predictive analytics on sports plays Upserve Restaurant table mgmt & POS for forecasting customer traffic TuSimple Computer Vision for Autonomous Driving Clarifai Computer Vision APIs

13 算法深度学习的到来编程模型 AWS 数据 GPUs 加速器

14 行业知识对数据科学家的要求计算编程数理方法

15 Amazon AI 基于深度学习的人工智能服务

16 Amazon AI 介绍 Apache MXNet 深度学习引擎

17 Amazon AI 介绍 Apache MXNet Polly 深度学习引擎文字转语音服务

18 Amazon AI 介绍 Apache MXNet Polly Rekognition 深度学习引擎语音服务图像分析

19 Amazon AI 介绍 Apache MXNet Polly Rekognition Lex 深度学习引擎语音服务图像分析语音识别和自然语义理解对话引擎

20 Amazon AI 介绍 Apache MXNet Polly Rekognition Lex 深度学习引擎语音服务图像分析对话引擎

21 Apache MXNet 可编程轻便高性能语法简单, 支持多种语言为移动端及 IoT 提供高性能模型在成百上千个 GPUs 之间近乎线性缩放

22 为什么使用 Apache MXNet? 基本开源在 AWS 上表现最佳纳入 Apache 孵化项目在 AWS 上对深度学习进行了优化 ( 与 AWS 集成整合 )

23 Apache MXNet 是亚马逊选择的深度学习框架

24 P2 实例, 深度系统映像, 及深度模板 P2 实例接近 4 万个 CUDA 核深度系统镜像 AMI 为深度学习预设置深度模板深度学习集群

25 AWS 深度学习系统镜像 (Amazon Linux, Ubuntu) 接近 4 万个 CUDA 核 Apache MXNet TensorFlow Theano Caffe Torch 预设置 CUDA 驱动 Anaconda, Python3 一键使用深度学习 + CloudFormation 模板 + 容器镜像

26 Amazon AI 介绍 Apache MXNet Polly Rekognition Lex 深度学习引擎语音服务图像分析对话引擎

27 Amazon Polly: 栩栩如生的语音服务将文本转换成生动的语音完全托管 47 种声音 24 种语言低延迟, 实时

28 我们先来听一听

29 Polly: 聚焦音质和发音 1. 自动且准确的文本处理过程 Today in Seattle, WA, it s 11 F "We live for the music" live from the Madison Square Garden.

30 Polly: 聚焦音质和发音 1. 自动且准确的文本处理过程 2. 智能且易于理解 Peter Piper picked a peck of pickled peppers.

31 Polly: 聚焦音质和发音 1. 自动且准确的文本处理过程 2. 智能且易于理解 3. 为文本赋予语义 Richard s number is Richard s number is 电话号码

32 Polly: 聚焦音质和发音 1. 自动且准确的文本处理过程 2. 智能且易于理解 3. 为文本赋予语义 4. 定制式发音 My daughter's name is Kaja. My daughter's name is Kaja.

33 Amazon Polly: 栩栩如生的语音服务具备一流深度学习深层功能易于使用且完全集成为产品而定制价格低廉框架的高质量

34 Amazon AI 介绍 Apache MXNet Polly Rekognition Lex 深度学习引擎文本转语音图像分析对话引擎

35 Rekognition: 搜索并理解视觉内容实时图像对象及场景检测人脸检测人脸搜索面部分析批处理分析

36 Amazon Rekognition 基于深度学习的图像识别物品场景识别面部分析面部比较面部确认

Rekognition: 对象及场景检测 Couch Furniture Bowl Entertainment Center Flower Arrangement Plant Vase Floral Design Flower Coffee Table Table Indoors Interior Design Room

37 Rekognition: 对象及场景检测 Couch Furniture Bowl Entertainment Center Flower Arrangement Plant Vase Floral Design Flower Coffee Table Table Indoors Interior Design Room Dining Table Apartment Housing Living Room Cabinet Fruit 97.7% 97.7% 94% 93.2% 86.4% 86.4% 86.4% 70.6% 70.6% 69.3% 69.3% 66.4% 66.4% 66.4% 58.9% 58.7% 58.7% 52% 51.9% 50.7%

Rekognition: 人脸检测 looks like a face appears to be male age range smiling appears to be happy not wearing eyeglasses not wearing sunglasses eyes

38 Rekognition: 人脸检测 looks like a face appears to be male age range smiling appears to be happy not wearing eyeglasses not wearing sunglasses eyes are open mouth is open does not have a mustache does not have a beard 99.9% 99.9% years old 98.3% 95.9% 99.2% 99.8% 73.2% 72.1% 87.4% 97.4%

39 Rekognition: 人脸辨认

40 } "contentstring": { "Attributes": [ "ALL" ], "TargetImage": { "Bytes": "..." }, "SourceImage": { "Bytes": "..." }, "SimilarityThreshold": 0 } { "FaceMatches": [ { "Face": { "BoundingBox": { "Height": , "Left": , "Top": , "Width": }, "Confidence": }, "Similarity": 78 }, { "Face": { "BoundingBox": { "Height": , "Left": , "Top": , "Width": }, "Confidence": }, "Similarity": 0 } ], "SourceImageFace": { "BoundingBox": { "Height": , "Left": 0.376{ "FaceMatches": [ { "Face": { "BoundingBox": { "Height": , "Left": , "Top": , "Width": }, "Confidence": }, "Similarity": 78 }, { "Face": { "BoundingBox": { "Height": , "Left": , "Top": , "Width": }, "Confidence": }, "Similarity": 0 } ], "SourceImageFace": { "BoundingBox": { "Height": , "Left": , "Top": , "Width": }, "Confidence": } } , "Top": , "Width": }, "Confidence": } }

41 Rekognition: 搜索并理解视觉内容具备一流深度学习深层功能易于使用且完全集成为产品而定制价格低廉框架的高质量

42 Amazon AI 介绍 Apache MXNet Polly Rekognition Lex 深度学习引擎文本转语音图像分析对话引擎

43 会话式交互的出现第一代 : 面向机器的交互打孔机记存器第二代 : 面向控制及翻译指针滑杆第三代 : 面向意愿对话界面

44 Lex: 在语音与文本之间建立自然会话式交互语音及文本高效在移动设备网页及使用 Slack 与 Messenger 企业连接器聊天机器人 Alexa 其他设备上的语音交互进行文本交互 ( 其他功能即将推出 ) Salesforce Microsoft Dynamics Marketo Zendesk Quickbooks Hubspot

45 Flight Booking Origin Destination Departure Date

46 Book a flight to Shanghai Flight Booking Origin Destination Departure Date

47 Book Flight Shanghai Book a flight to Shanghai Automatic Speech Recognition Flight Booking Origin Destination Departure Date

48 Book Flight Flight booking Knowledge Graph Shanghai Shanghai Hongqiao Book a flight to Automatic Natural Language Grammar Shanghai Speech Recognition Understanding Graph Utterances Flight Booking Origin Destination Departure Date

49 Book Flight Flight booking Knowledge Graph Shanghai Shanghai Hongqiao Book a flight to Automatic Natural Language Grammar Shanghai Speech Recognition Understanding Graph Utterances Flight Booking Origin Destination Shanghai Hongqiao Departure Date

50 Book Flight Flight booking Knowledge Graph Shanghai Shanghai Hongqiao Book a flight to Automatic Natural Language Grammar Shanghai Speech Recognition Understanding Graph Location Location Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date

51 Book Flight Flight booking Knowledge Graph Shanghai Shanghai Hongqiao Book a flight to Automatic Natural Language Grammar Shanghai Speech Recognition Understanding Graph Location Location Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date When would you like to fly? Prompt

52 Book Flight Flight booking Knowledge Graph Shanghai Shanghai Hongqiao Book a flight to Automatic Natural Language Grammar Shanghai Speech Recognition Understanding Graph Location Location Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date When would you like to fly? Polly When would you like to fly? Prompt

53 Next Friday Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date When would you like to fly?

54 Next Friday Next Friday Automatic Speech Recognition Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date

55 Flight booking Knowledge Next Friday 5/16/2017 Graph Next Friday Automatic Speech Recognition Natural Language Understanding Grammar Graph Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date 5/16/2017

56 Flight booking Knowledge Next Friday 5/16/2017 Graph Next Friday Automatic Speech Recognition Natural Language Understanding Grammar Graph Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date 5/16/2017

57 Flight booking Knowledge Next Friday 5/16/2017 Graph Next Friday Automatic Speech Recognition Natural Language Understanding Grammar Graph Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date 5/16/2017 Your flight is booked for next Friday Confirmation

58 Flight booking Knowledge Next Friday 11/18/2016 Graph Next Friday Automatic Speech Recognition Natural Language Understanding Grammar Graph Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date 5/16/2017 Your flight is booked for next Friday Polly Your flight is booked for next Friday Confirmation

59 Flight booking Knowledge Next Friday 5/16/2017 Graph Next Friday Automatic Speech Recognition Natural Language Understanding Grammar Graph Utterances Flight Booking Origin Destination Beijing Shanghai Hongqiao Departure Date 5/16/2017 Hotel Booking

60 开发者面临的挑战语音识别身份确认安全语义理解通信平台可用独立数据处理缩放业务逻辑测试移动应用传统语音对话界面需要结合大量复杂的算法和技术

61 Lex: 在语音与文本之间建立自然会话式交互具备一流深度学习深层功能易于使用且完全集成为产品而定制价格低廉框架的高质量

62 Echo 利用 Alexa Skills Kit 开发 Alexa! 用户 Alexa 开发者

63 Demo: Echo Dot Ask Questions Connect Bluetooth devices Control music Control smart home devices Enable skills Weather updates Get sport scores Find a place Listen to audible books Shopping Tell me a joke

64 Amazon AI: 接下来的发展方向? Apache MXNet Polly Rekognition Lex 新的 APIs 深度学习引擎语音服务图像分析对话引擎和工具

65 亚马逊人工智能 :AWS AI 应用开发员数据科学家 Amazon Rekognition 图像识别 & 分析 Amazon Polly 文本 - 语音转换 Amazon Lex 自然语言理解 (NLU) & 自动语音识别 (ASR) AWS Deep Learning AMI 快速轻松的使用可扩展的多种深度学习框架 Amazon Machine Learning 托管的机器学习服务 P2 AWS P2 Instance 通用 GPU 实例

66 深度学习应用场景云服务个人助理服务器监控金融服务自动语音识别自然语音处理金融预测分析智能投顾医疗卫生图像视频标签推荐系统汽车交通定向广告人脸识别网络侵入检测欺诈侦测

67 Amazon AI: 开始使用吧

68 谢谢

PowerPoint Presentation

PowerPoint Presentation 使用 Amazon Rekognition 实时分析图像王元恺,AWS 解决方案架构师沟通交流你好沟通交流你好 Amazon Rekognition Image 采用深度学习技术的完全托管式图像识别服务搜索, 鉴别并组织百万的图片对象和场景检测面部分析面部比较面部识别名人识别图像审核文字识别对象和场景检测 DetectLabels { }, { }, { },... "Confidence":