コンピュータビジョンを支える深層学習技術の新潮流

コンピュータビジョン最新技術から、AutoMLやモデルコンパイラ、推論チップなど、開発・運用周りのトレンドをご紹介

Transcript

1. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Masaki Samejima Machine Learning Solutions Architect, Amazon Web Services Japan. 2019.2.14 ؝ٝؾُ٦ةؽآّٝ׾佄ִ׷ 帾㾴㷕统䪮遭ך倜惐崧 Developers Summit 2019
2. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Agenda • ؝ٝؾُ٦ةؽآّٝך⹛ぢ • ؝ٝؾُ٦ةؽآّٝךꟚ涪䪮遭 • ؝ٝؾُ٦ةؽآّٝך麊欽䪮遭 • ؝ٝؾُ٦ةؽآّٝ׾佄ִ׷ع٦سؐؑ،ך㾜Ꟛ
3. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝך⹛ぢ
4. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝהכ • ➂꟦כ歗⫷ַ׵㢳ֻך䞔㜠׾䖤׷ֿהָדֹ׷ • ؝ٝؾُ٦ةח׮ず圫ח➂꟦ך鋔鋙׾䭯׋ׇ״ֲהׅ׷ 䪮遭ָ؝ٝؾُ٦ةؽآّٝ Demographic Data Facial Landmarks Sentiment Expressed Image Quality General Attributes 겣ⴓ匿ך⢽ ➂꟦כ겣׾鋅ג邌䞔 זו׾铣׫《׸׷
5. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝחֶֽ׷帾㾴㷕统ך娖〷 2012 歗⫷钠陎 SuperVision[1] ILSVRC2012⮚⹧ [1] A. Krizhevsky, et al., Imagenet classification with deep convolutional neural networks, NIPS 2012. [2] R Girshick, et al., Rich feature hierarchies for accurate object detection and semantic segmentation, CVPR 2014. [3] I.J. Goodfellow, et al., Generative Adversarial Nets, NIPS 2014. [4] V. Badrinarayanan, et al, SegNet: A Deep Convolutional Encoder-Decoder Architecture for Image Segmentation. PAMI 2017 2014 暟⡤嗚⳿ R-CNN[2]ָPascal VOCד礵䏝ぢ♳ 歗⫷欰䧭 侯㼎涸欰䧭طح زٙ٦ؙGAN[3] إًؚٝذ٦ءّٝ ؾؙإٕ⽃⡘ךٓك ؚٔٝSegNet[4] 2015
6. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 굲鬨涸ז䚍腉ぢ♳
   歗⫷钠陎 https://gluon-cv.mxnet.io/model_zoo/classification.html senet_154 resnet_v1d resnet_v1c resnet_v1b resnet_v1 densenet darknet VGG resnet_v2 mobilenet mobilenetv2 0.80 0.75 0.70 Accuracy 1000 2000 䱿锷鸞䏝 #sample/sec. 3000 4000 • ImageNetח㼎׃ג80%ך礵䏝 • V100 GPUד侧⼪卐ך䱿锷鸞䏝 ⰼך㣐ֹׁכًٌٔ؟؎ؤ
7. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 굲鬨涸ז䚍腉ぢ♳
   暟⡤嗚⳿ https://gluon-cv.mxnet.io/model_zoo/detection.html mAP 10 100 䱿锷鸞䏝 #sample/sec. ⰼך㣐ֹׁכ ًٌٔ؟؎ؤ 40 35 30 yolo3 faster_rcnn ssd • 嗚⳿걄㚖ָ姻鍑ה鵚ְ(IoUָ㣐ְֹ)暟⡤ חאְגծmAP礵䏝ָ30-40%玎䏝 • 鸞䏝꬗דכ굲鬨涸חぢ♳
8. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 굲鬨涸ז䚍腉ぢ♳
   إًؚٝذ٦ءّٝ https://gluon-cv.mxnet.io/model_zoo/segmentation.html 0 10 20 30 40 50 60 70 80 90 100 fcn_resnet101 psp_resnet101 deeplab_resnet101 fcn_resnet101 psp_resnet101 deeplab_resnet101 deeplab_resnet152 COCO ر٦ةإحز VOC ر٦ةإحز IoU 暟⡤׀הח姻׃ֻ陎ⴽדֹ׋걄㚖ךⶴさָぢ♳
9. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ״׶넝䏝ז钠陎堣腉 3如⯋ך暟⡤嗚⳿ [1] [1] B. Tekin, et al., Real-Time Seamless Single Shot 6D Object Pose Prediction, CVPR 2018. [2] R. Girdhar, et al., Detect-and-Track: Efficient Pose Estimation in Videos, CVPR 2018. [3] L. Chen, et al., MaskLab: Instance Segmentation by Refining Object Detection with Semantic and Direction Features, CVPR 2018. ⹛歗ך暟⡤嗚⳿٥鷄騊 [2] 暟⡤׀הך إًؚٝذ٦ءّٝ[3]
10. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. GANך黝欽걄㚖ך䭁㣐 歗⫷ך欰䧭 Noise 歗⫷ך㢌䳔 Text-to-image [3] (and Image-to-text) ظ؎ؤַ׵歗⫷פ 遹僤歗⫷ַ׵㖑㔳פ[1] 歗⫷ך넝鍑⫷䏝⻉[2] [1] P. Isola, et al., Image-to-Image Translation with Conditional Adversarial Nets, CVPR 2017. [2] C. Ledig, et al., Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, CVPR 2017. [3] S. Reed, et al., Generative Adversarial Text to Image Synthesis, ICML 2016.
11. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 歗⫷ַ׵圫ղז؝ٝذؗأزך钠陎 Saliency (岣湡걄㚖) ך嗚⳿ [1] 鋔简倯ぢ׾嗚⳿ׅ׷鄲縧ד ⡲䧭׃׋ر٦ةإحز׾ⵃ欽 [1] N. Liu, et al., PiCANet: Learning Pixel-wise Contextual Attention for Saliency Detection, CVPR 2018. [2] Z. Li, et al., MegaDepth: Learning Single-View Depth Prediction from Internet Photos, CVPR 2018. 歗⫷ך㤴遤䱿㹀 [2] ְ׹׿ז錬䏝ַ׵乆䕦׃׋歗⫷ד 㤴遤ֹ׾䱿㹀׃גر٦ةإحز׾⡲䧭
12. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 㼰ꆀر٦ةח㼎ׅ׷䮋䨌 0 2 4 6 8 10 12 14 16 18 20 1 2 3 4 5 6 7 8 9 1011121314151617181920 겣歗⫷ر٦ةإحز 卐侧 歗⫷ID ⴓ겲 㼎韋 ⴓ겲㼎韋כ Ӎ ח⡂גְ׷ [1] O. Vinyals, et al., Matching Networks for One Shot Learning, arXiv:1606.04080 ر٦ةכ䌢ח⼧ⴓח ֮׷׻ֽדכזְ • ⦐ղ׾陎ⴽׅ׷ٌرٕ׾⡲׷ֿהכ㔭ꨇ • ⡂גְ׷ַוֲַדⴻⴽׅ׷[1] 嫰 鯰 㼎 韋
13. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 侯㼎涸歗⫷ך欰䧭ה꣇䖴 • Deep Learningךٌرٕכ䠐㢩חظ؎ؤח䓲ְ • ػٝتך歗⫷חظ؎ؤ׾⸇ִ׷הذشؖؠٕהⴻ㹀ׁ׸׷⢽ X. Yuan, et al., Adversarial Examples: Attacks and Defenses for Deep Learning, IEEE Trans Neural Netw Learn Syst. 2019.
14. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝךꟚ涪䪮遭
15. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝךꟚ涪ך铬겗 • 넝䚍腉ז،ٕ؞ٔؤيך㹋鄲ךꨇ׃ׁ • 帾㾴㷕统ؿٖ٦يٙ٦ؙך✉甧 • ⸬桦涸ז؝٦سך㹋鄲٥رغحؚך㔭ꨇׁ • ر٦ةⳢ椚٥⸇䊨ծٌرٕ锃侭⡲噟ך頾䬐 ع؎ٖكٕ㹋鄲׾㹋植 ׅ׷ؿٖ٦يٙ٦ؙ ONNXח״׷ؿٖ٦ي ٙ٦ؙ꟦ךٌرٕ㢌䳔 AutoML Define-by-run ח״׷ فؚٗٓىؚٝ
16. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ع؎ٖكٕ㹋鄲׾㹋植ׅ׷ؿٖ٦يٙ٦ؙ • 剑⯓畭ךٌرٕך㹋鄲װ㷕统幥׫ٌرٕ׾ꂁ䋒 • ِ٦ؠכٌرٕ׾ٗ٦س׃גծⱄ㷕统ծ䱿锷ָ〳腉 TensorFlow models TF slim GluonCV ChainerCV PyTorchCV
17. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ResNet㹋鄲ך嫰鯰 (Gluon vs MXNet) num_unit = len(units) assert(num_unit == num_stages) data = mx.sym.Variable(name='data') if dtype == 'float32': data = mx.sym.identity(data=data, name='id') else: if dtype == 'float16': data = mx.sym.Cast(data=data, dtype=np.float16) data = mx.sym.BatchNorm(data=data, fix_gamma=True, eps=2e-5, momentum=bn_mom, name='bn_data') (nchannel, height, width) = image_shape if height <= 32: # such as cifar10 body = mx.sym.Convolution(data=data, num_filter=filter_list[0], kernel=(3, 3), stride=(1,1), pad=(1, 1), no_bias=True, name="conv0", workspace=workspace) else: # often expected to be 224 such as imagenet body = mx.sym.Convolution(data=data, num_filter=filter_list[0], kernel=(7, 7), stride=(2,2), pad=(3, 3), no_bias=True, name="conv0", workspace=workspace) body = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn0') body = mx.sym.Activation(data=body, act_type='relu', name='relu0') body = mx.sym.Pooling(data=body, kernel=(3, 3), stride=(2,2), pad=(1,1), pool_type='max') for i in range(num_stages): body = residual_unit(body, filter_list[i+1], (1 if i==0 else 2, 1 if i==0 else 2), False, name='stage%d_unit%d' % (i + 1, 1), bottle_neck=bottle_neck, workspace=workspace, memonger=memonger) for j in range(units[i]-1): body = residual_unit(body, filter_list[i+1], (1,1), True, name='stage%d_unit%d' % (i + 1, j + 2), bottle_neck=bottle_neck, workspace=workspace, memonger=memonger) bn1 = mx.sym.BatchNorm(data=body, fix_gamma=False, eps=2e-5, momentum=bn_mom, name='bn1') relu1 = mx.sym.Activation(data=bn1, act_type='relu', name='relu1') MXNet 㹋鄲♧鿇 from mxnet.gluon.model_zoo import vision resnet18 = vision.resnet18_v1() Gluon 㹋鄲
18. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ONNX (Open Neural Network Exchange) MXNet Caffe2 PyTorch TF CNTK CoreML Tensor RT NGraph SNPE • ぐ珏ؿٖ٦يٙ٦ؙד圓眠׃׋ ٌرٕ׾ONNXח㢌䳔ծONNX ַ׵ⴽךؿٖ٦يٙ٦ؙח㢌䳔 〳腉 • ؔ٦فٝا٦أד֮׶ծ؝ىُ صذ؍ח״׏גꟚ涪ׁ׸גְ׷
19. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ONNX ךⰻ鿇圓䧭 Protocol Buffers ךغ؎شٔؿ؋؎ٕ • ؝ٝػؙزד⸬桦涸 • ֮׵ײ׷فٓحزؿؓ٦يד⹛⡲ • APIח״׷铣׫鴥׫٥剅ֹ鴥׫ָ〳 腉 Protocol Buffers Graph Operator Tensor, … Operator Definitions ONNX Python API
20. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Define-and-run ה Define-by-run • Define-and-run • طحزٙ٦ؙ׾㹀纏׃גַ׵ر٦ة׾Ⰵ⸂ׅ׷倯䒭 • TensorFlow, MXNetָⴱ劍ַ׵㼪Ⰵ • Define-by-run • طحزٙ٦ؙ׾㹀纏׃זָ׵ر٦ة׾Ⰵ⸂ׅ׷倯䒭 • Chainerָ䱰欽ծ׉ך䖓ծPyTorch, TensorFlow, MXNetפה岚⿹
21. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Define-and-run ה Define-by-run ך⢽ Define-and-run Define-by-run def our_function(A, B): C = A + B return C A = Load_Data_A() B = Load_Data_B() result = our_function(A, B) A = placeholder() B = placeholder() C = A + B our_function = compile(inputs=[A, B], outputs =[C]) A = Load_Data_A() B = Load_Data_B() result = our_function(A, B) https://gluon.mxnet.io/chapter07_distributed-learning/hybridize.html Define Run Define, Run
22. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Define-by-run ךًٔحز Define-and-run Define-by-run def our_function(A, B): C = A + B return C A = Load_Data_A() B = Load_Data_B() result = our_function(A, B) A = placeholder() B = placeholder() C = A + B our_function = compile(inputs=[A, B], outputs =[C]) A = Load_Data_A() B = Load_Data_B() result = our_function(A, B) ر٦ة׾䟝㹀׃׋鏣鎘 ָ䗳銲㢌⻉ׅ׷ر٦ ة׾䪔ְב׵ְ 㢌⻉חさ׻ׇ׋鏣鎘٥رغح ָؚ〳腉
23. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AutoML הכ • 堣唒㷕统׾㹋欽⻉ׅ׷׋׭חכ圫ղזةأָؙ㶷㖈 • ر٦ة鼅ⴽ٥欰䧭暴䗙䬄⳿ٌرٕ鼅䫛, etc. D. Bayor, et al., TFX: A TensorFlow-Based Production-Scale Machine Learning Platform, KDD 2017.
24. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AutoML הכ • ꬊ㼔Ꟍ㹺ך׋׭ח䗳銲זةأؙ׾荈⹛⻉ׅ׷ AutoML ך灇瑔ָ鹌׬ • ICML 2014ַ׵ AutoML ךٙ٦ؙءّحف*ָأة٦ز • ⚺銲ז灇瑔زؾحؙ • ع؎ػ٦ػًٓ٦ةך剑黝⻉ 堣唒㷕统חְֶגծ➂䩛ח״׷鏣㹀ָ䗳銲ז ع؎ػ٦ػًٓ٦ة׾⸬桦״ֻ䱱稊 • Meta-Learning, Learning to learn 堣唒㷕统דⵃ欽ׁ׸׷剑黝⻉䩛岀זו׾ر٦ة ַ׵㷕统ׅ׷倯岀㷕统ׅ׷倯岀׾㷕统ׅ׷ * https://sites.google.com/site/automlwsicml14/
25. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AutoML ָ׮׋׵ׅ堣唒㷕统ءأذي 荈⹛דر٦ة　꧊٥侭⪒ 荈⹛ד剑黝ז堣唒㷕统׾㹋遤 ِ٦ؠכ堣唒㷕统ך،ٕ؞ٔؤي׮ر٦ةך侭⪒倯岀׮ 孡חׅ׷䗳銲ָזְ
26. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. AutoML ך⢽
    Amazon Forecast User CSV file 1. S3חر٦ة،حفٗ٦س 2. Forecast חر٦ةإحز ה✮庠،ٕ؞ٔؤيך鏣㹀 3. Forecast ָ✮庠ٌرٕ圓眠 4. ٌرٕח״׷儗禸⴨ر٦ة ך✮庠ז׵ןח〳鋔⻉ ِ٦ؠחה׏ג剑⡚ꣲ䗳銲זֿהכծر٦ةך،حفٗ٦سהⴱ劍鏣㹀
27. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝך麊欽䪮遭
28. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝך麊欽ך铬겗 • 㷕统׃׋ٌرٕד⸬桦״ֻ䱿锷ׅ׷׋׭ך 橆㞮׾㺁僒ח圓眠׃׋ְ • ٌرٕח״׷䱿锷穠卓ך㧅䔲䚍׾鐰⣣דֹ זְ Model Server ٌرٕ؝ٝػ؎ٓ Interpretable ML
29. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Model Server ٌرٕرفٗ؎ך铬겗 • 堣唒㷕统ٌرٕח㼎׃גծ䱿锷ؙٔؒأز׾Ⳣ椚ׅ׷橆㞮ָ䗳銲 • 䱿锷橆㞮ך圓眠如痥דכծٖ؎ذٝءָ㉏겗חז׷ֿה׮֮׷ Model Server ך䕵ⶴ • ⸬桦ך葺ְ䱿锷橆㞮׾㺁僒ח圓眠〳 • REST/RPCח״׷؎ٝة٦ؿؑ٦أדؙٓ؎،ٝزַ׵ⵃ欽׃װְׅ 堣唒㷕统ٌرٕ Model Server Mobile client Deploy REST/RPC
30. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. TensorFlow Serving [1] C. Olston, et al., TensorFlow-Serving: Flexible, High-Performance ML Serving, NIPS 2017. • Controller, Synchronizer׾穗גծ䱿锷欽ךServing job ׾醱侧欰䧭 • Routerכծ䱿锷ؙٔؒأز׾「ֽ׷הծServing job ח鷏⥋
31. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. MXNet Model Server https://aws.amazon.com/jp/blogs/news/model-server-for-apache-mxnet-v1-0-released/ • ٌرٕ涫ꐮծ䱿锷ؙٔ ؒأزך鷏⥋׾REST API䕎䒭ד遤ִ׷ • MMS 1.0ד䚍腉ָぢ♳ ׃ծず儗1,000،ؙإأ ד׮ⰋגⳢ椚〳腉 MMS 1.0 MMS 0.4
32. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 帾㾴㷕统ٌرٕ؝ٝػ؎ٓ • 帾㾴㷕统ךٌرٕ׾㹋遤ׅ׷ع٦س ؐؑ،ח䘔ׄגٌرٕ׾㢌䳔׃ծٌ رٕך䱿锷鸞䏝׾ぢ♳ • ؝ٝػ؎ٕ䖓ךٌرٕכ鯪ꆀזٓٝ ة؎يד㹋遤〳腉דծ؟؎ؤך㣐ֹ ְ帾㾴㷕统ؿٖ٦يٙ٦ָؙ♶銲 • ؝ٝػ؎ٓך⢽ • AWS, SageMaker Neo • Nvidia, TensorRT Raspberry Piך䱿锷儗꟦ ResNet18 Mobilenet 11.5x 2.2x
33. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. SageMaker Neo / TVM ח״׷剑黝⻉ • Operator Fusion 怴皾׾מהתה׭ח׃ג넝鸞⻉ • Data Layout Transformation 4x4ך遤⴨怴皾׾遤ֲ㜥さכծر٦ة׮4x4 ךة؎ٕ朐חׅ׷ • Tensor Expression and Schedule Space ⸬桦⻉ׅ׷׋׭ח⡚ٖكٕז邌植ח㢌䳔ծ怴皾׾أ؛آُ٦ؚٔٝ • Nested Parallelism with Cooperation ぐأٖحسָر٦ة׾⼿⸂׃ג《׶ח遤ֹծⰟ剣ًٌٔחְֶגְֻ • etc… T. Chen, et al., TVM: An Automated End-to-End Optimizing Compiler for Deep Learning, OSDI 2018.
34. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. TensorRT ח״׷剑黝⻉ • Layer & Tensor Fusion 醱侧ךٖ؎َח״׷Ⳣ椚׾1אךٖ؎َךⳢ椚חתה׭׷ • FP16 and INT8 Precision Calibration FP32ך➿׻׶חFP16װINT8׾ⵃ欽׃ג؝ٝػؙزח׃ծ怴皾׾넝鸞⻉ • Kernel Auto-Tuning ثُ٦صׁؚٝ׸׋ؕ٦طٕ׾荈⹛鼅䫛׃ծ⢽ִלծ殴鴥׫ך怴皾׾剑黝⻉ • Dynamic Tensor Memory ًٌٔךⵃ欽׾幾׵׃׋׶ׅ׷ֿהדծًٌٔⶴ׶䔲גךؔ٦غقحس׾⵴ꤐ • Multi Stream Execution 醱侧ךⰅ⸂أزٔ٦يד⚛⴨חⳢ椚׃ג䱿锷〳腉 https://devblogs.nvidia.com/tensorrt-3-faster-tensorflow-inference/
35. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Interpretable ML: ⽃秪זٌرٕ ءٝفٕז堣唒㷕统ٌرٕדכծٌرٕ׾鍑匿׃גծוךر٦ةָ✮ 庠װⴓ겲ח剣⸬ַ׾濼׷ֿהָדֹ׷⢽: 寸㹀加ծSVMծGBTկ C. Molnar, Interpretable Machine Learning, https://christophm.github.io/interpretable-ml-book/ ➂〡 >900♰ ➂〡 < 900♰ ꬗琎 < 2000 km2 匌❨鿪 㣐ꢻ䏍 ꬗琎 > 2000 km2 ❨鿪䏍 ⢽ִלծ寸㹀加ך㜥さծ ♳ךظ٦سךקֲָꅾ銲 דծֿך㜥さծ➂〡ָⴓ 겲ךꅾ銲ז䩛ַָ׶
36. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Interpretable ML for computer vision • 帾㾴㷕统ך״ֲח醱꧟⻉׃׋ٌرٕדכծٌرٕך湫䱸涸ז鍑匿כ 㔭ꨇկ؝ٝؾُ٦ةؽآّٝך㜥さכծ歗⫷׾Ⰵ⸂׃ג䮙⹛׾鍑匿կ • 歗⫷ך♧鿇׾妀衅ׇׁג׮姻׃ֻ钠陎דֹ׸לծ婍׏׋鿇ⴓכ钠陎 חְֶגꅾ銲ז皘䨽،ٝؕ٦ה׫זׇ׷կ ⯋歗⫷ ،ٝؕ٦ ،ٝؕ٦⟃㢩׾何㢌׃ג׮姻׃ֻ钠陎 M.T. Ribeiro, et al., Anchors: High-Precision Model-Agnostic Explanations, AAAI 2018.
37. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. ؝ٝؾُ٦ةؽآّٝ׾佄ִ׷ ع٦سؐؑ،ך㾜Ꟛ
38. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 䱿锷欽ثحف • ⡚؝أزד넝鸞ח⹛⡲ׅ׷䱿锷ثحفךꟚ涪 • 㷕统欽ثحفהכ欽鷿ָ殯ז׷ • 㷕统欽כ侧⼧卐ך歗⫷׾鋅חغحثה׃ג㷕统 • 䱿锷欽כ1卐1卐ךِ٦ؠךؙٔؒأز • 涪邌ׁ׸גְ׷ثحفך⢽ • AWS Inferentia • Intel Nervana
39. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. Machine Learning on FPGA • 넝⸬桦זFPGA׾⢪欽׃׋帾㾴㷕统 • AWS F1 instanceדכծꟚ涪خ٦ٕ׾ろ׭׋ Amazon Machine Image׾ⵃ欽דֹ׷ • 殴׫鴥׫صُ٦ٕٓطحزٙ٦ؙך״ֲח粸׶鵤׃鎘皾ָ䗳銲 ז׮ךכLoop tiling ח״׷剑黝⻉ָ䗳銲 [1] [1] C. Zhang, et al., Optimizing FPGA-based Accelerator Design for Deep Convolutional Neural Networks, FPGA 2015.
40. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 堣唒㷕统ؒحآرغ؎أ • طحزٙ٦ؙ♶銲דծ⡚ٖ؎ ذٝءד䱿锷ָ〳腉 • 㼭㘗ךⰻ詿GPUד׮帾㾴㷕统 ך䱿锷ָ〳腉ח
41. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. 堣唒㷕统ؒحآرغ؎أ
42. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. תה׭
43. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. תה׭ • ؝ٝؾُ٦ةؽآّٝך灇瑔כծ礵䏝٥鸞䏝ךぢ♳׌ֽדזֻծ ،فٔ؛٦ءّٝך䌴׮䎢ָ׏גְ׷կ • 帾㾴㷕统،فٔ؛٦ءّٝךꟚ涪٥麊欽ך⸬桦⻉ָ岣湡ׁ׸גְ ׷կAutoMLהְ׏׋荈⹛⻉חꟼׅ׷灇瑔׮遤׻׸ծ➙䖓כAIך 孖⚺⻉ָ鹌׬կ • 帾㾴㷕统ח״׷؝ٝؾُ٦ةؽآّٝ׮ծ䱿锷׾㹋⚅歲ח䘔欽ׅ ׷ؿؑ٦ؤחկ䱿锷ثحفծؒحآرغ؎أָ➙䖓崞鬨կ
44. © 2019, Amazon Web Services, Inc. or its Affiliates. All rights reserved. إحءّٝך׀䠐鋅ծ׀䠬䟝׾ֶ耀ַׇֻ׌ְׁ https://amzn.to/aws_dev