아마존의 딥러닝 기술 활용 사례

아마존의 딥러닝 기술 활용 사례

아마존의 딥러닝 기술 활용 사례 - 윤석찬 (AWS 테크니컬 에반젤리스트) from Amazon Web Services Korea

아마존의 딥러닝 기술 활용 사례 - 윤석찬 (AWS 테크니컬 에반젤리스트)

1. © 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved. @ A ,

2. . • : : : : N • : I LP MF D N G • & A , : : • , : : • :

3. . .

4. ) ( • , • • 5 1 &&

5. ( 5 ) • E ) • C -- ( ) - ) • ) ) ) ) 9 13 % 5

6. - ( 2 • , , 3 P B D • S I W M - ) 1, 0

7. 1) 0 (/ • , • ' 2 1 5 5

8. 43 , 2 1 • 5 6 : 7 2 • % % • 0 1 % • 5 0 43 ) 2 1 (

9. Transforming Industrial Processes with Deep Learning (MAC301), AWS re:Invent 2016 https://www.youtube.com/watch?v=AHUaor0odh4

10. ArrivalImage Tower ( ) ArrivalImage Tower Departure Image Tower ArrivalImage Tower Departure Image Tower

11. -

12. - • • • Krizhevsky’s CNN CIFAR CNN Best Hand- Engineered Model

13. - Original image Activation map Binarymap 2.0 1.0 Google Net Conv Conv (3*3) Avg Pool 3*3 1024 channels

14. : • 3 S A . / . - . / / .

15. - 2 7 • • 1 ) ( 0 6

16. - 12 -.0 , 2 • , • 7 ( ) 8 ( https://www.amazon.com/b?node=16008589011

17. Active Customers Up Nearly 5X Tens of Millions of Alexa-Enabled Devices

18. ,0 0 + Alexa Voice Service + 5 2 Alexa Skills Kit

19. https://github.com/alexa/alexa-avs- sample-app/wiki/Raspberry-Pi https://echosim.io

20. Deep Learning in Alexa (MAC202), AWS re:Invent 2016 https://www.youtube.com/watch?v=TYRckcVm4WE

21. S A 8 B 2 0 M3 S E Corpus size 20K+ hours GPUs - g2.2xlarge B A G P U C B S Distributed SGD

22. 0 100,000 200,000 300,000 400,000 500,000 600,000 0 10 20 30 40 50 60 70 80 Framespersecond Number of GPU workers DNN training speed Strom, Nikko. "Scalable Distributed DNN Training using Commodity GPU Cloud Computing." INTERSPEECH. Vol. 7. 2015.

23. 1 4.75 8.5 12.25 16 1 4.75 8.5 12.25 16 Speedup(x) # GPUs Resnet 152 Inceptin V3 Alexnet Ideal P2.16xlarge (8 Nvidia Tesla K80 - 16 GPUs) Synchronous SGD (Stochastic Gradient Descent) 91% Efficiency 88% Efficiency 16x P2.16xlarge by AWS CloudFormation Mounted on Amazon EFS # GPUs

24. ## train num_gpus = 4 gpus = [mx.gpu(i) for i in range(num_gpus)] model = mx.model.FeedForward( ctx = gpus, symbol = softmax, num_round = 20, learning_rate = 0.01, momentum = 0.9, wd = 0.00001) model.fit(X = train, eval_data = val, batch_end_callback = mx.callback.Speedometer(batch_size=batch_size))

25. http://gluon.mxnet.io - • ,W NTca I • ( P C W d MS K H b • ) A ) A A A X • A ,C C X NEW!

26. • A Kumar, et al, Just ASK: Building an Architecture for Extensible Self-Service Spoken Language Understanding, https://arxiv.org/abs/1711.00549 • R Maas, et al, Domain-Specific Utterance End-Point Detection for Speech Recognition - Proc. Interspeech 2017, http://www.isca-speech.org/archive/Interspeech_2017/pdfs/1673.PDF • B King et al, Robust Speech Recognition Via Anchor Word Representations - Proc. Interspeech 2017, http://www.isca-speech.org/archive/Interspeech_2017/pdfs/1570.PDF • A Kumar et al, Zero-shot learning across heterogeneous overlapping domains - Proc. Interspeech 2017, http://www.isca-speech.org/archive/Interspeech_2017/pdfs/0516.PDF • M Sun et al, Max-pooling loss training of long short-term memory networks for small-footprint keyword spotting, Spoken Language Technology Workshop (SLT), 2016 IEEE • F Ladhak et al, LatticeRnn: Recurrent Neural Networks Over Lattices - Proc. Interspeech 2016, http://www.isca- speech.org/archive/Interspeech_2016/pdfs/1583.PDF • S Panchapagesan et al, Multi-Task Learning and Weighted Cross-Entropy for DNN-Based Keyword Spotting - Proc. Interspeech 2016, http://www.isca-speech.org/archive/Interspeech_2016/pdfs/1485.PDF • R Maas et al, Anchored Speech Detection - Proc. Interspeech 2016, http://www.isca- speech.org/archive/Interspeech_2016/pdfs/1346.PDF • M Sun et al, Model Shrinking for Embedded Keyword Spotting, 2015 IEEE 14th International Conference on Machine Learning and Applications (ICMLA) • N Strom, Scalable distributed DNN training using commodity GPU cloud computing, Annual Conference of the International Speech Communication Association 2015, http://www.isca- speech.org/archive/interspeech_2015/papers/i15_1488.pdf

27. NEW! “Alexa, start the meeting.” “Alexa, dial 555-8000.” “Alexa, lower the blinds.” “Alexa, ask Salesforce which big deals closed today.”

28. 44.1% 7.7% 3.0% 2.3% 1.0% 1.4% 0.7% 2.2% 0.5% 0.9% 4 ) 0 2 1 % 37 % ( 2 8

29. 2012 2013 2015 20172014 20162008 2009 2010 2011 516 24 48 61 82 159 280 722 1,017 LAUNCHES 1,300+

30. Most robust, fully featured technology infrastructure platform

31. - - FRAMEWORKS AND INTERFACES AWS DEEP LEARNING AMI Apache MXNet TensorFlowCaffe2 Torch KerasCNTK PyTorch GluonTheano PLATFORM SERVICES VISION AWS DeepLensAmazon SageMaker LANGUAGE Amazon Rekognition Amazon Polly Amazon Lex Amazon Rekognition Video Amazon Transcribe Amazon Comprehend Alexa for Business VR/AR Amazon Sumerian APPLICATION SERVICES Amazon Machine Learning Amazon EMR & SparkMechanical Turk INSTANCES GPU (G2/P2/P3) CPU (C5) FPGA (F1) Amazon Translate

32. F R A M E W O R K S A N D I N T E R FA C E S NVIDIA Tesla V100 GPUs P3 1 Petaflop of compute NVLink 2.0 5,120 Tensor cores 128GB of memory ~14X faster than P2 P3 Instance Deep Learning AMI Frameworks PLATFORM SERVICES VISION LANGUAGE VR/IR APPLICATION SERVICE AWS DeepLensAmazon SageMaker Amazon Machine Learning Amazon EMR & SparkMechanical Turk AWS DEEP LEARNING AMI Apache MXNet TensorFlowCaffe2 Torch KerasCNTK PyTorch GluonTheano INSTANCES GPU (G2/P2/P3) CPU (C5) FPGA (F1)

33. 2 0 3 p3.2xlarge = $5 per hour p3.2xlarge x 20 = $100 per hour ) ( 1 20

34. Spot Instances (75% ↓) = $30 per hour

35. 3 $aws ec2-run-instances ami-b232d0db --instance-count 20 --instance-type p3.2xlarge --region us-east-1 $aws ec2-stop-instances i-10a64379 i-10a64280 ...

36. CUSTOMERS RUNNING MACHINE LEARNING ON AWS TODAY

37. ( ) !

38. H J . 31 N 31 , - N 31 2 , - , - - NEW!

39. FRAMEWORKS AND INTERFACES AWS DEEP LEARNING AMI Apache MXNet TensorFlowCaffe2 Torch KerasCNTK PyTorch GluonTheano PLATFORM SERVICES VISION AWS DeepLensAmazon SageMaker LANGUAGE Amazon Rekognition Amazon Polly Amazon Lex Amazon Rekognition Video Amazon Transcribe Amazon Comprehend Alexa for Business VR/AR Amazon Sumerian APPLICATION SERVICES Amazon Machine Learning Amazon EMR & SparkMechanical Turk INSTANCES GPU (G2/P2/P3) CPU (C5) FPGA (F1) Amazon Translate

40. C A D ,65 .88 387 9 ,41 g g 2 8 a g C 55 ES 2 8 re t D J t M Ip i J D L J n 2 8 g g ,65 y a 2 8 D D W L J n 2 + 2 2 2 H D t t A u H Discrete Classification, Regression Linear Learner Supervised XGBoost Algorithm Supervised Discrete Recommendations Factorization Machines Supervised Image Classification Image Classification Algorithm Supervised, CNN Neural Machine Translation Sequence to Sequence Supervised, seq2seq Time-series Prediction DeepAR Supervised, RNN Discrete Groupings K-Means Algorithm Unsupervised Dimensionality Reduction PCA (Principal Component Analysis) Unsupervised Topic Determination Latent Dirichlet Allocation (LDA) Unsupervised Neural Topic Model (NTM) Unsupervised, Neural Network Based

41. CA “With Amazon SageMaker, we can accelerate our Artificial Intelligence initiatives at scale by building and deploying our algorithms on the platform. We will create novel large-scale machine learning and AI algorithms and deploy them on this platform to solve complex problems that can power prosperity for our customers." - Ashok Srivastava, Chief Data Officer, Intuit

42. Mdt h z r bg S Yo z 2 U k$ nw c a$ aW w ( e s s aW p LS 0C K 7 5 B c 097 4 C m 10 MIN NEW! HD video camera Custom-designed deep learning inference engine Micro-SD Mini-HDMI USB USB Reset Audio out Power • Intel Atom Processor • Intel Gen9 graphics • Ubuntu OS- 16.04 LTS • 100 GFLOPS performance • Dual band Wi-Fi • 8 GB RAM • 16 GB Storage (eMMC) • 32 GB SD card n P ) . A / C K C 1 ,: 23 • 4 MP camera with MJPEG • H.264 encoding at 1080p resolution • 2 USB ports • Micro HDMI • Audio out • AWS Greengrass • clDNN Optimized for MXNet

43. FRAMEWORKS AND INTERFACES AWS DEEP LEARNING AMI Apache MXNet TensorFlowCaffe2 Torch KerasCNTK PyTorch GluonTheano PLATFORM SERVICES AWS DeepLensAmazon SageMaker Amazon Machine Learning Amazon EMR & SparkMechanical Turk INSTANCES GPU (G2/P2/P3) CPU (C5) FPGA (F1) VISION LANGUAGE Amazon Rekognition Image Amazon Polly Amazon Lex Amazon Rekognition Video Amazon Transcribe Amazon Comprehend Alexa for Business VR/AR Amazon Sumerian APPLICATION SERVICES Amazon Translate

44. • L B A M 2 • ,

45. , ?

46. , 2 .4 4 3 3 1

47. 43 2, 43 1 .

48. ) 4A d ) 4A m I f W TRg TRg M a n o e ck i L à i lb TRg P o S (1 2 352 ( 2 ( 2 A ( 2 C (1 ( 2 2A ( 2 2 2A ( 2 2 4 3 ( 2 2 2

49. AWS ML Customers APPLICATION SERVICES Amazon Lex Amazon Polly Amazon Comprehend Amazon Translate Amazon Transcribe Amazon Rekognition Image Amazon Rekognition Video PLATFORM SERVICES Amazon SageMaker AWS DeepLens FRAMEWORKS AND INTERFACES AWS Deep Learning AMI Apache MXNet Caffe2 CNTK PyTorch TensorFlow Theano Torch Gluon Keras AWS ML Platform DATA LAKE STORAGE Amazon S3 SECURITY Access Control Encryption COMPUTE Powerful GPU and CPU Instances ANALYTICS Amazon Athena Amazon Redshift and Redshift Spectrum Amazon EMR (Spark, Hive, Presto, Pig) AWS Glue Amazon Kinesis Amazon QuickSight Amazon Macie AWS Organizations AWS Cloud Platform

50. 1 1 7 • FC S TF ITTQS BWS BNBZP DPN LP NBDI F MFB • 1FFQ 6FB .7 ITTQS BWS BNBZP DPN LP NBDI F MFB BN S • 7?8FT ITTQS BWS BNBZP DPN LP NX FT • F SP 2MPW ITTQS BWS BNBZP DPN LP TF SP GMPW 1 7 017 • . F F T 7BDI F 6FB 7 0P • ITTQS WWWYPUTUCF DPN QMBYM ST-M ST 96I AQ ZULFX 8D K /CN K U QU • . F F T 7BDI F 6FB FSS P S • ITTQS WWWYPUTUCF DPN QMBYM ST-M ST 96I AQ ZULF =0IA QL 8WQI:N 7 1 7 21 1 1 • FC S TF ITTQS WWWB G P T F S DPN • M FS ITTQ WWWSM FSIB F FT . 2 P T F S Q FSF TBT P S

51. © 2018, Amazon Web Services, Inc. or its Affiliates. All rights reserved.