诸葛越 - 算法无处不在

苏凡白

2017/12/18 发布于 技术 分类

Hulu 是一家国际领先的视频平台,提供优质电影,电视剧点播节目,和ESPN, FOX, CNN, HBO 等各大电视网络的直播节目,在美国有数千万付费用户。Hulu 技术架构最为先进的一点是人工智能和机器学习算法的广泛应用:在个性化内容推荐,搜索,视频内容理解,视频传输和播放,广告预测和定向,安全检测,决策支持,甚至视频编辑和客服系统。 机器学习算法的背后是专门打造的大规模数据处理系统。“算法无处不在”,是Hulu当今和未来技术架构的定位。在这个主题演讲里,我会给大家介绍如何从基础研究到系统架构来建造这样一个先进的互联网平台,以及业务层面的内容。

文字内容
1. 算法无处不在 Make Algorithms Ubiquitous Hulu Architecture: Today and Tomorrow 诸葛越 Vice President, Hulu December 8, 2017
3. 诸葛越 Vice President, Hulu 工作经历 •  Hulu全球副总裁,北京研发中心总经理 •  Landscape Mobile联合创始人,CEO •  雅虎北京全球研发中心产品总监 •  微软北京研发中心资深项目经理 •  Senior Data Architect, Yahoo US 教育背景 •  PhD, MS Stanford University Computer Science •  MS Stony Brook University Applied Math •  清华大学计算机科学与技术系
4. •  Machine Learning Algorithms in Industry – 机器学习算法企业应用场景 •  ML Algorithms in Hulu –无处不在的机器学习算法在Hulu •  Democratize ML In Hulu – Hulu ML算法案例研究 •  AI Platform In Hulu – Hulu人工智能平台 •  Future Opportunities – 展望未来
5. Background – 行业背景 AI is everywhere ML centric business face recognition self-driving car medical image analysis … ? consumer business online entertainment online game travel / tourist application e-commerce …
6. ML Algorithm – 机器学习算法的技术定义 Program vs Algorithm vs a collection of instructions the logic to solve a class of problems Machine Learning Algorithm the method to learn the logic DATA + MODEL
7. Trend – ML算法应用的趋势 targeting recommendation prediction specific scenarios 特定场景 垂直 => 水平 特型 => 通用 专家 => 平民化 ubiquitous 无处不在
8. •  Machine Learning Algorithms in Industry – 机器学习算法企业应用场景 •  ML Algorithms in Hulu –无处不在的机器学习算法在Hulu •  Democratize ML In Hulu – Hulu ML算法案例研究 •  AI Platform In Hulu – Hulu人工智能平台 •  Future Opportunities – 展望未来
9. Hulu The Premier Digital Video Company Best Quality Content Video-On-Demand and Live Broadcasting
10. ML Algorithms in Hulu ML算法在Hulu的应用 Viewers! Advertising! Content!
11. Algorithms for Viewers 面向用户的算法 cover story smart start recommendation reason layout management
12. Algorithms for Viewers 面向用户的算法 onboarding cold start convert-to-pay … Home Page cover story smart start recommendation reason layout management … ads auto-play QoS to QoE …
13. Algorithms for Content and Advertising 面向内容和广告的算法 Content Advertising content valuation video understanding content based video encoding … rich targeting inventory prediction …
14. •  Machine Learning Algorithms in Industry – 机器学习算法企业应用场景 •  ML Algorithms in Hulu –无处不在的机器学习算法在Hulu •  Democratize ML In Hulu – Hulu ML 算法案例研究 •  垂直 – Reco •  水平 – Video Content Understanding •  水平到垂直 – Contextual + Ads •  AI Platform In Hulu – Hulu人工智能平台 •  Future Opportunities – 展望未来
15. Recommendations: StoreShelf •  Item-based Collaborative Filtering •  Matrix Factorization •  Deep Neural Network
16. DNN-based Relevance Algorithm
17. Auto-Play, Time Series Model Search Watch Search Browse New Show? Browse Try it out Browse Watch Binge Watch Browse Watch Search Watch Watch Watch Watch Watch Watch Watch Recurrent Neural Network Timeline
18. Video Content Understanding: From Inside 从内部理解视频内容 Video derived meta-data 元数据抽取和发现 Breaks •  Ads break, end credit, ... Arts •  Video summary, ... Ceremonyҁ0.91҂ Wedding ҁ0.82҂ Brideҁ0.72҂ Eventҁ0.38҂ Groomҁ0.36҂ Womanҁ0.31҂ Mealҁ0.28҂ Marriageҁ0.27҂ Dressҁ0.25҂ Wedding receptionҁ0.18҂ Objects •  Actor, car, animal, ... Events •  Cooking, wedding, …
19. Video Content Understanding: From Inside 从内部理解视频内容 Live Channel Thumbnail Preview 直播频道预览
20. Contextual Ads: 视频和广告的结合 Ads Provider Hulu Content •  Deliver semantic related or visual similar ads to viewers during playback Labels of Target Scenes Scene-Level Auto-Tagging •  视觉连贯的广告 Matching •  语义相关的广告 •  Use video derived data to match ads and content Pairs Ads Delivery score:0.86
21. Contextual Ads: 视频和广告的结合 •  视觉连贯的广告 •  语义相关的广告
22. Contextual Ads: 视频和广告的结合 •  视觉连贯的广告 •  语义相关的广告
23. •  Machine Learning Algorithms in Industry – 机器学习算法企业应用场景 •  ML Algorithms in Hulu –无处不在的机器学习算法在Hulu •  Democratize ML In Hulu – Hulu ML算法案例研究 •  AI Platform In Hulu – Hulu人工智能平台 •  Future Opportunities – 展望未来
24. AI Platform – 人工智能平台 Application Model Feature Data Infrastructure All ML Applications Application Specific ML Models Shared ML Models Shared Features User Data Static / Dynamic Content Data Audio / Video / Text, Metadata Ads Data Audio / Video / Text Network / DevOps / Big Data GPU Cluster Frame Storage
25. AI Platform – 人工智能平台 Application Onboarding Convert-to-Pay Cold Start Auto-Play Recommendation Content Valuation QoS to QoE Ad Targeting …… Model Application Specific ML Models User Conversion Model User Engagement Model Image Understanding …… Feature Data Infrastructure User Persona User Embedding Knowledge Graph Content Segments Content Embedding User Data Static / Dynamic Content Data Audio / Video / Text, Metadata Ads Data Audio / Video / Text Network / DevOps / Big Data GPU Cluster Frame Storage
26. Computational Mode – 计算模式 model Offline ~days/hours Nearline ~minutes data Online ~milliseconds
27. •  Machine Learning Algorithms in Industry – 机器学习算法企业应用场景 •  ML Algorithms in Hulu –无处不在的机器学习算法在Hulu •  Democratize ML In Hulu – Hulu ML算法案例研究 •  Hulu AI Platform – Hulu人工智能平台 •  Future Opportunities – 展望未来
28. Future Opportunities – 展望未来 •  Smarter interactions 更有创意的交互体验 •  Live Video 实时的个性化体验 •  Personalized design 个性化设计