UPLOAD

    788

    Machine Learning Applications using Amazon SageMaker - 16 September- 11:00

    Published: October 13, 2019

    AWS Loft IStanbul 2019 Machine Learning Applications using Amazon SageMaker 16 September- 11:00

    Comments

    Machine Learning Applications using Amazon SageMaker - 16 September- 11:00

    • 1. Amazon SageMaker © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Amazon SageMaker Build, Train, and Deploy Machine Learning Models Quickly & Easily, at scale Başar Kızıldere Account Manager @ AWS
    • 2. Slide220 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark ML FRAMEWORKS & INFRASTRUCTURE The Amazon ML stack: Broadest & deepest set of capabilities AI SERVICES Vision | Documents | Speech | Language | Chatbots | Forecasting | Recommendations ML SERVICES Data labeling | Pre-built algorithms & notebooks | One-click training and deployment Build, train, and deploy machine learning models fast Easily add intelligence to applications without machine learning skills Flexibility & choice, highest-performing infrastructure Support for ML frameworks | Compute options purpose-built for ML
    • 3. Slide1443 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark ML FRAMEWORKS & INFRASTRUCTURE The Amazon ML stack: Broadest & deepest set of capabilities AI SERVICES REKOGNITION IMAGE POLLY TRANSCRIBE TRANSLATE LEX REKOGNITION VIDEO Vision Speech Chatbots ML SERVICES Frameworks Interfaces Infrastructure EC2 P3 & P3dn EC2 C5 FPGAs GREENGRASS ELASTIC INFERENCE Language Forecasting Recommendations TEXTRACT New COMPREHEND & COMPREHEND MEDICAL New New New FORECAST PERSONALIZE GROUND TRUTH New NOTEBOOKS AWS MARKETPLACE New ALGORITHMS REINFORCEMENT LEARNING New TRAINING OPTIMIZATION (NEO) New DEPLOYMENT HOSTING New AMAZON SAGEMAKER
    • 4. Slide1455 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Collect and prepare training data Choose and optimize your ML algorithm Train and Tune ML Models Set up and manage environments for training Deploy models in production Scale and manage the production environment Pre-built notebooks for common problems Amazon SageMaker: Build, Train, and Deploy ML Models at Scale Built-in, high performance algorithms One-click training on the highest performing infrastructure Model Optimization One-click Deployment Fully managed with auto-scaling for 75% less
    • 5. Successful models require high-quality data © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Successful models require high-quality data
    • 6. Slide1460 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Raw Data Human Annotations Automatic Annotations Training Data Human Annotations Amazon SageMaker Ground Truth Label machine learning training data easily and accurately
    • 7. AWS Marketplace for Machine LearningML algorithms and models available instantly © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark AWS Marketplace for Machine Learning ML algorithms and models available instantly Subscribe in a single click Available in Amazon SageMaker KEY FEATURES Automatic labeling via machine learning IP protection Automated billing and metering Browse or search AWS Marketplace SELLERS Broad selection of paid, free, and open-source algorithms and models Data protection Discoverable on your AWS bill BUYERS
    • 8. Amazon SageMaker Neo: Train once, run anywhere © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Amazon SageMaker Neo: Train once, run anywhere Neo
    • 9. Amazon SageMaker NeoTrain once, run anywhere with 2x the performance © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Amazon SageMaker Neo Train once, run anywhere with 2x the performance Automatic optimization Broad framework support KEY FEATURES Get accuracy and performance Open-source device runtime and compiler, 1/10th the size of original frameworks Broad hardware support
    • 10. What is a Reinforcement Learning (RL) environment? © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark What is a Reinforcement Learning (RL) environment? Representation of the real world Programmed to represent real- world conditions Enables interaction with user or a computer program Dynamic and updates itself based on the interactions and programmed behavior
    • 11. Amazon SageMaker RLReinforcement learning for every developer and data scientist © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Amazon SageMaker RL Reinforcement learning for every developer and data scientist Broad support for frameworks Broad support for simulation environments 2D & 3D physics environments and OpenGym support Support Amazon Sumerian, AWS RoboMaker and the open source Robotics Operating System (ROS) project Fully managed Example notebooks and tutorials KEY FEATURES
    • 12. Slide1270 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Predictions drive complexity and cost in production Inference (Prediction) 90% Training 10%
    • 13. Amazon Elastic Inference © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Amazon Elastic Inference Lower inference costs Match capacity to demand Available between 1 to 32 TFLOPS per accelerator KEY FEATURES Reduce deep learning inference costs up to 75%
    • 14. Recap: Amazon SageMaker © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Recap: Amazon SageMaker Data labeling & Pre-built notebooks for common problems Model and algorithm marketplace & Built-in, high performance algorithms One-click training on the highest performing infrastructure Train once, run anywhere & Model Optimization One-click deployment Collect and prepare training data Choose and optimize your ML algorithm Train and tune models Set up and manage environments for training Deploy models in production Scale and manage the production environment Fully managed with auto-scaling for 75% less Amazon EC2 P3 Instances Amazon SageMaker RL Amazon SageMaker Ground Truth Amazon Elastic Inference AWS Marketplace for Machine Learning Amazon SageMaker Neo
    • 15. Slide1463 © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Amazon Confidential and Trademark Thank You!
    • 16. Slide1 AWS SageMaker ile Makine Öğrenmesi Uygulamaları Günay Özkan, ML & AI Team Lead @Armut 16 Eylül 2019
    • 17. Slide2 Armut Ne Yapar? Erişim Güven Kalite Görünürl ük Erişim Hizmet Alan Hizmet Veren
    • 18. Slide3 Temel Problemler
    • 19. Slide4 Armut AWS Veri Altyapısı
    • 20. Slide5 AWS/SageMaker Üzerinde Çalışan Projeler 1.Hizmet Tavsiyesi N Hizmet alan bir kullanıcının alacağı N+1’inci hizmeti tahmin etme. Eğitme sıklığı: Haftada bir defa. Çalışma sıklığı: On Demand. 2. Sahte yorum tahmini Bir işe yapılan yorumun HV’nin kendisi veya bir yakını tarafından girilmiş olma ihtimalinin tahmin edilmesi. Eğitme sıklığı: Ayda bir defa. Çalışma sıklığı: Günde bir defa.
    • 21. Slide6 AWS/SageMaker Üzerinde Çalışan Projeler 3. Amazon Rekognition Sisteme yüklenen fotoğrafların denetimi 4. Dolandırıcılık tespiti Sisteme kayıt olan hizmet verenlerin dolandırıcı olma ihtimalinin tespiti. Eğitme sıklığı: Ayda bir defa. Çalışma sıklığı: Günde bir defa. 5. HA-HV Eşleştirme Gelen işlerin en uygun hizmet verenlerle eşleştirilmesi Eğitme sıklığı: Haftada bir defa. Çalışma sıklığı: On demand.
    • 22. Slide7 Hizmet Tavsiyesi
    • 23. Slide8 Hizmet Tavsiyesi - Modeller ve Algoritmalar 1.A Priori Association Rule Mining. Birlikte alınan hizmetlerin analizi. ['Kır Düğünü', 'Dış Çekim Fotoğraf', 'Düğün Salonu'] 2. SVD Matris Çarpanlara Ayırma ile boyut azaltma ve eksik elemanların tahmini. 3. ARIMA Zaman Serisi Analizi
    • 24. Slide9 Hizmet Tavsiyesi - Modeller ve Algoritmalar 4. Hizmet2Vec Hizmet embedding Tarihsel hizmet alım verisi üzerinden hizmet türlerinin yüksek boyutlu uzaya gömülmesi. 5. Neural Network Hizmet alım verisindeki non lineer örüntüleri yakalayan yapay sinir ağı modeli.
    • 25. Slide10 HA - HV Eşleştirme Ciddiyet Lokasyon Hizmet alım geçmişi Takvim Lokasyon İş tercihleri Hizmet Alan Hizmet Veren İş İş Bilgileri Aranma izni Tahmini fiyat ve kalite
    • 26. Slide11 Eşleştirme - Modeller ve Algoritmalar 1. Logistic Regression İşin kazanılma ihtimali, yani kalitesinin belirlenmesi için ilk olarak Logistic Regression denendi. 2. Random Forest Yine işin kalitesinin tahmini için Random Forest daha iyi sonuç verdi. 3. Gradient Boosting Hem işin kalitesinin tahmini, hem de hizmet verenlerle eşleştirilmesi için en iyi sonuçlar Gradient Boosting ile elde edildi.
    • 27. Slide12 Sahte Yorum Belirleme - Modeller 1.Random Forest 20 Kadar feature üzerinden supervised olarak eğitiliyor. %98.6 Precision, %83 Recall Logistic Regression, Tek Karar Ağacı ve Random Forest denendi, en iyi sonuçlar Random Forest ile elde edildi.
    • 28. Slide13 Profil Fotoğrafı Analizi 1.Profil fotoğrafları ile Hizmet Veren kalitesi arasında ilişki bulma
    • 29. Slide14 SageMaker Süreçleri 1.Notebook’lar Çeşitli problemler için geliştiren özel modeller notebook’larda tutuluyor. 2. Eğitme Modeller özel docker image’ları içerisinde eğitilerek sonuçlar S3’e kaydediliyor. 3. Optimizasyon Gerekirse hyper-parameter optimizasyonu SageMaker üzerinde hızlı bir şekilde gerçekleştiriliyor 4. Deployment S3’teki modeller SageMaker üzerinde endpoint olarak çalıştırılıyor 5. A/B Testing Kullanıcıların farklı alt kümelerine farklı modellerin açılması, modeller arası canlı performans karşılaştırması.
    • 30. Slide15 Yakın Gelecekte 1.Mevcut Modellerin iyileştirilmesi Daha fazla NLP, Zaman serisi modelleme, Kullanıcı Bilgisi. 2. Yeni Modellerin Geliştirilmesi Armut’un çözmeye çalıştığı problemlerin daha fazlası için ML tabanlı modellerin geliştirilmesi. Örn. Marketing LTV ve Churn Tahmini, HV rating tahmini, Kullanıcı şikayet email’lerinin cevaplarla veya çözüm personeli ile otomatik eşleştirilmesi.
    • 31. Slide16 Teşekkürler! Sorular? Armut Engineering https://twitter.com/ArmutEng