![- 2018 9 11 .
AI vs 瑚 蟆一 螳!
瑚 讌讌 = 碁螳 企螻 2016 蟾讌.
瑚 讌蠍 = 2016~2020(or 2022?)
瑚 讌 = 2020? 2025? (螳 覓語)
[] AI螳 瑚 覈 覿朱ゼ . 讀, 轟 覿殊襷 企!
[] But,瑚覲企 AI螳 覿手 襷讌 蟆企,
[讌] 企 覿手 螻 蠍 企, 蠍壱 りり .](https://image.slidesharecdn.com/ssgmachinelearningjourneyhoondongkim201809-181002220309/85/E-commerce-BigData-Scale-AI-Journey-12-320.jpg)
E-commerce BigData Scale AI Journey
- 1. SSG.COM Machine Learning(BigData & AI) Journey Hoon Dong Kim BigData&AI / SSG.COM hoondongkim@emart.com
- 2. I am 蟾 Chief Partner 瑚 蠏碁９ 殊 SSG.COM BigData&AI Part Leader れ覈(Korea Spark User Group) 伎讌 BigData 覿 Microsoft MVP(Most Valuable Professional) 2016, 2017 AI 覿 Microsoft MVP(Most Valuable Professional) 2018 http://hoondongkim.blogspot.kr https://www.facebook.com/kim.hoondong
- 4. れ願蠍一
- 5. 2008 2015 2006 20XX(???) 覿朱 °螳 襴螻, 譴蟲覓朱螳 螳覦覃? 襴覦覦 豐覦一?
- 6. 2017 (Walmart vs Amazon vs Korean Player) 譬 觜 麹 豬渚 蟆曙???
- 7. 永姻看鉛看乙顎艶(襷)
- 8. 螻手碓 2015 prologue (襷) 襷ろ ( 瑚 vs AI ) 企語 語
- 9. 企語語 Deep Learning 蠍一 - 旧 ? ( By Amazon ) れ暑譯曙 覲伎企 讌 覓手唄れ 襷譟伎 螻襴讀 蟲蟆 覦一 蟆朱, 讌 れ螳 覦一 覓手唄れ 蟆 谿場譴. 谿所 襷 ? Long Tail れ?
- 10. 螻手碓 2016,2017 - prologue (襷) - AlphaGo & 觚 襷ろ ( 瑚 vs AI ) れ 覿殊 瑚 願鍵蠍 覃伎 一螻螳 譯朱蠍 . 覦 伎碁 9 vs 螻(蟲蠍) 1 : 4 ロ伎 vs 觚(ETRI 蟲 AI) 350 : 510 讀1 一轟, 讀2 一轟, レ 襷, 蟆 襦蠏碁 一轟
- 11. 覦 Reinforcement Learning 蠍一 - 旧 ? (By Alibaba & Taobao) 螳誤 讀螳 覦朱 蟆 伎 轟 れ螳朱 By Alibaba & Taobao https://arxiv.org/abs/1803.00710 れ螳奄 螳誤 Training & Serving
- 12. - 2018 9 11 . AI vs 瑚 蟆一 螳! 瑚 讌讌 = 碁螳 企螻 2016 蟾讌. 瑚 讌蠍 = 2016~2020(or 2022?) 瑚 讌 = 2020? 2025? (螳 覓語) [] AI螳 瑚 覈 覿朱ゼ . 讀, 轟 覿殊襷 企! [] But,瑚覲企 AI螳 覿手 襷讌 蟆企, [讌] 企 覿手 螻 蠍 企, 蠍壱 りり .
- 13. Our Goal!
- 14. 磯Μ螳 覈 蟆も
- 15. 襷譟 螳誤 蟾** 襦蠏語 覦** 襦蠏語 襷譟 豌 伎 螳 れ 螳 螳誤 螻, 企 語 Item 螳誤 . 蟯螻 豺 蟯螻 譬襯 る.
- 16. Youtube 螳誤 蟾** 襦蠏語 覦** 襦蠏語 觚 豌 伎 螳 れ 螳 螳誤 螻, 企 語 Item 螳誤 .
- 17. Our Strategy!
- 18. AI 襯 螻 蠍一 譴 Google 覲企ゼ 一 危ｴ覲伎.
- 19. Google Did! 蟲蠍 覲 譯殊 覦 伎 Google 覦 蠍一 OpenSource 讌 蠍一 2003 GFS Hadoop HDFS 2004 MapReduce Hadoop MapReduce 2006 Chubby Zookeeper 2006 BigTable HBase 2010 Pregel Neo4J, Spark Graph-X 2010 Dremel Spark 2011 Tenzing Hive, Spark SQL 2012 Spanner, F1 RDB Sharding + Kafka + Redis : RDB 覿 Scale Out 覦 蠍壱 ろ 2013 Omega Docker(Mesos) 2014 Word2Vec , DataFlow れ Word2Vec (NLP 蠍一) , Beam れ 2015 Tensorflow , Borg れ Tensorflow (Deep Learning Framework), kubernetes れ 2016 AlphaGo, DeepQN, WaveNet れ 覲牛覓語襯 願屋 れ Deep Learning 覦覯襦 2017 ~ PathNet, Transformer れ Istio, knative , spinnaker れ Deep Learning 旧 讌, 覲企 螻谿 覦覯 れ AI serving れ 譯朱 覦 K8S Ecosystem も BigData 蠍一 NoSQL 蠍一 覿 RDB 蠍一 MicroService & DevOps & Container& Data Workflow 蠍一 Deep Learning & AI 蠍一 & K8S Ecosystems Deep Learning 蠍一 伎 BigData, NoSQL, MicroService 煙 Core 蠍 蠍磯 襾殊 螳豢.
- 20. AI 覦! 1. BigData 蠍一 煙 (2003~2010) 2. NoSQL 蠍一 煙 (2007~2013) 3. Container & Microservice 蠍一 煙 (2014~) 4. Deep Learning 螻 蠍一 煙 (2015~) 蠏碁Μ螻 5. 蠍一れ 牛 讌 !
- 21. We Did! (BigData Scale Computing)
- 22. We Did! (RealTime Lambda ろ豌)
- 23. 殊 螻襯 覦螳螻 ! 螻手碓襯 覿 . (BigData Eco System Infra) 碁 襦蠏碁ゼ 蠍企. 觜一危 讌 覿 誤朱ゼ 襷. 觜一危 覦一襦 螻手碓 螻 覿 螻 螳襯 . 覦 . (RealTime Layer / kafka , Spark Streaming / ELK) れ螳朱 一危 ろ碁殊 覿. FDS, 覲伎蟯, 覈磯 讀螳朱 豌 襯 ル朱 觜. 覩碁襯 豸 . (Mining / Machine Learning / Deep Learning R&D) 螻螳 蟯 螳螻 螻 螻 蟆 螳 蟆 豢豌. 覩碁 讌 蟯螻 覦 豈 一 覲企 ROI 蟆 覦磯. 覦譯朱ゼ 豸″. 豕 碁 蟆暑襯 豸″. 螳蟆 襴伎 襷讌 朱 語狩讌 豕 螳蟆 豸″. 覩碁 豸′ 螻 . (Machine Learning / Deep Learning Production) Chatbot 一 , 企語 -> 蟆 覦 豢豌 螻 NLP , Item2Vec, etc 覩碁 豸′ 螳誤 覦 れ螳煙 . (BigData Scale Deep Learning) れ 襯 螳誤 Deep Learning 觜. BigData Scale Training , BigData Scale Inference. Auto Scale Out 覈 覦壱, 覓伎讌 Deep Learning 覈語 讌 覦 覦壱. 蠍一ヾ Machine Learning , Deep Learning 覈語 freshness , personalization 螻. 5~6 3~5 1~2 2~3 讀 螳 螻, 螻螳 螻, 蠍一 蠍一襦 覿 讌螳 錫.
- 24. Production A/B Test 譴 蟆! Sample Data, Selected Feature 蟲 覈 譯1 or 1 覈蟆 蟲 覈 豕螻 VS 襷 Data(or Data) Simple 覈 豕 (襷 螳 or 譴れ螳) 螳誤 覈 企 觜襯願, れ蟆, れ螳朱 朱 , Kaggle 殊!
- 25. 襴狩 豢豌(蟲譟) + 襴狩 覿 By Netflix Lambda Architecture Netflix 100襷 Competition 一麹 螻襴讀 讌 螳?
- 26. We Did! (BigData Scale AI) Public cloud Hadoop / NoSQL Spark / Anaconda Spark ML Mahout Sk-learn/gensim Tensorflow/CNTK TensorflowOnSpark Keras On-Premise Spark On-premise Serverless / K8SServerless PaaS & AI Infra Microservices Web/Was Docker / K8S 螳譬 PaaS AI BaaS/SaaS PaaS App Container PaaS GPU Container ML/DL PaaS & BigData Lake PaaS & GPU Notebooks 4~5 2~3 豕蠏
- 27. Deep Learning Inference (On Docker Microservice) 焔レ http://hoondongkim.blogspot.kr/2017/12/deep-learning-inference-serving.html VM(for minibatch size 1 inference) Docker(for minibatch size 1 inference)
- 28. Our Action!
- 29. 稃磯碁覺,譯狩,覓伎語拘,覓伎語
- 30. 覓朱 蠏碁Μ螻, Data Driven
- 31. *** 豕 覈 (By BigData Driven + AI) Human Model1(2016, 糾蠍磯) Model2(2017豐, 觚) Model3(2017襷,BigData + 觚) Model3(2018,BigData + 觚 + AI) Accuracy 60% 90% 碁ル豌 NN覈 0覈 10覈 More Data + More Freshness Only More Data
- 32. ***豸 覈 (By BigData + AI) Item Average Accuracy 68% 90% 49% 38% Regression / Python Decision Tree / Python 豐蠍 覦 / 15 貉 4螳 / 15 貉 6螳 / 15 Random forest / Python 貉 6螳 / 15 Time Series Deep Learning + TensorflowOnSpark 貉 26螳 / 3 Random forest /Spark ML 貉 6螳 / 3 Rule Python BigData + ML XGBoost / Python 貉 6螳 / 15 3 一危. Incremental Learning. Time Series Deep Learning Tensorflow + Spark (Scale Out) Single Host + DL BigData + DL 貉 26螳 / 15 Time Series Deep Learning / Tensorlfow Data 蠍 /覲 Data 蠍 /覲
- 33. 貉狩 AI & CS覺 瑚 蠏碁９ 殊 SSG.COM
- 34. ル 企語 蟆
- 35. ル慨蠍 覺
- 36. SSG 覺 AI 覈 ろ豌 Top Level classifier Model #4 Sequential Classifier Model #5 , 殊, CS, Domain Word Embedding Model #1 襦る 襦る 襦る 襦る 1 -> 1 / Class N Context Manager Model #6 NER Model #7 Word2Vec, GloVe, Swivel Flow Designer Rule Manager API Manager 語蟯襴 Tracking Log 豈 I/F API Handling Semantic Embedding Model #3 覈 蠏 OCR Model #2 蠍 Item2Vec, Doc2Vec Item2Vec Word2Vec 豺危螻襴 Tag Item 覈 Tag 朱 N -> 1 / Class M Semantic Search Rule Base CS Bot FAQ/QNA Pair Generative Bot Flow Designer Rule Manager API Manager 殊 Pair Extreme multiclass Wide Pair Model #9 Semantic Search Model #8 Seq2seq Model #10 襦る history Sequence 1 : 1 Pairs Doc2Vec, Graph Embedding
- 37. Intent Classifier (Model #4) 1. Word2Vec + CNN (Batch Normalize + Augmentation) 2. Word2Vec + LSTM 3. Word2Vec + CNN + LSTM 4. Word2Vec + Bidirectional GRU 5. Word2Vec + Bidirectional GRU + Attention Network 6. FastText 7. Glove + LSTM (BigDL on Spark Cluster) 72.30% 73.94% 72.97% 74.36% 73.15% 72.50% 75.25% 450螳 Multiple Class , Top 1 覓語. 8. Data 覦 Argumentation. 89.6% Data 殊. , 企Π讌, Argumentation, Data 覲 讌 觜 焔ロレ 覿覿 覈 語 蟆 13. Swivel + word Embedding +char Embedding + Hybrid LSTM + れ Approach 譟壱 豕譬 93.6% ML Model 1 Na誰ve Bayes ML Model 2 TF-IDF + SVM 48.26% 61.05% ML DL 襦 Model 焔 レ But, 碁 .
- 38. 蟆磯
- 39. 讌 螳覦 蟆 レ ! 讌 讌煙 ! 蟆讀 蟆れ MashUp! 旧蟆曙レ 蟆 觜襯願 蟆 譬. (Over Engineering 覦讌) (Open Source + Cloud PaaS) 讀 螳 Hot Develop 覦れ 豢蟲 蟆も No-Ops (or Dev Ops, Agile, Serverless, Microservices 願 覈 覃 . 訖.) Scale (壱 Scale Out, Scale Down, 一, レ) Low Cost (螳誤 豕, 豕 蠍磯 ) Performance ( 焔, 覈 焔, 螳覦一)
- 40. Low Cost 覦 蠏 覦覯 1 Docker , 1螳 朱? 1 朱? 1,000,000 Transaction 襷??? Tensorflow 豌企 蠍 蟾讌襷 蟲 螳 螻襦 る, Deep Learning Output 蠏碁 碁 螳 Graph DB , Hash DB 朱 覦 Index 覃, 螻 企 Engineering 覦 . 螻襦 覃, 襴覦 覦 Scale Front AI 觜 り 襷れ 觜襯願 襷れ 危蟆 螳. ル 覈碁 Graph DB, Microservice 襯 覈 襯願, NoSQL 螳覦 ル 覈襯願 Or, CloudML (on GCP) GPU VM > CPU VM > GPU Docker > K8S PaaS > Docker PaaS > Docker BaaS > Serverless Microservice GPU on-premise 1 GPU VM , 1 朱?
- 41. AI ! Think Big. 蠍瑚 覲願 Plan 語一. 轟レ 螻殊 一壱讌 襷. Bottom Up! ( Not Top Down! ) ｸ麹伎, 轟 覿れ 讌 伎螳 , る伎(Self Motivated )れ 蠍壱,螳覦 螻, 企 伎襯 螻ろ覃 讌 螻 襦 . 蟆 蠏碁Μ螻 觜襯願 . 蠍 る, 10 譴願, 2~3 讌 15 譴 讌襷, 1~2 れ 覃, 蟆曙覲企 . Mashup 螻, 蟆壱 Merge 螻, 覦覲 螳 れ. 蠏碁 Top 殊? Motivation. 蠍磯. 蟆曙＾. 覦レ 覿. Self Motivated 讌れ 蟆 蠍語 伎伎朱 蟆. 螳螳 襷 讌 蟆 . 覦譯殊 Data 覓語 覦 譯殊襯 磯殊♀鍵 螻襴讀 磯狩蠍 所, 豢 label Data 磯狩蠍 . 襦 伎覲企, 蠍磯狩 譴 朱, Cloud PaaS 蟆曙 Key 螳 .
- 42. Thank You 蠍壱 覓語 http://hoondongkim.blogspot.kr https://www.facebook.com/kim.hoondong Q & A