Share
Embed
小規模データセットに対する機械学習のアプローチ
1クラスあたりの画像枚数が少ないデータセットに対する機械学習の手法をサーベイしました。Data Engineering & Data Analysis Workshop #7 @CyberAgentでの発表。
Masaki Samejima
December 14, 2018
Transcript
34 slides
-
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Masaki Samejima Solutions Architect, Amazon Web Services Japan. 2018.12.14 㼭鋉垷歗⫷ر٦ةإحزח 㼎ׅ堣唒㷕统ך،فٗ٦ث Data Engineering & Data Analysis WS#7 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 荈䊹稱➜ 늦䃊 姻埠 ،وبٝ ؐؑـ ؟٦ؽأ آٍػٝ 堣唒㷕统أُ٦ءّٝ،٦ؗذؙز 㥨ֹזAWS؟٦ؽأ Amazon SageMaker, AWS Cloud9 馯 ٖٖؔؔSageMaker example Deep Learning Document ך缺鏬 https://s3.amazonaws.com/ja.gluon.mxnet.io/index.html https://github.com/harusametime/sagemaker-notebooks/ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 歗⫷ر٦ةإحزח㼎ׅ劍䖉 • 歗⫷ח㼎ׅ堣唒㷕统כծקר Deep Learning ח獳遤 • 钠陎׃ְぐؙٓأחծ⼧ⴓז歗⫷ָ֮ךָ椚䟝 0 2000 4000 6000 8000 0 1 2 3 4 5 6 7 8 9 MNIST Food-101 0 200 400 600 800 1000 1200 Sushi Ramen Edamame Takoyaki Gyoza … -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 植㹋ךر٦ةإحز 0 50 100 150 200 250 300 350 1 2 3 4 5 6 7 8 9 1011121314151617181920212223 0 2 4 6 8 10 12 14 16 18 20 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Labeled Faces in the Wild 겣歗⫷ر٦ةإحز Landmark Recognition Challenge ٓٝسو٦ؙ钠陎 https://www.kaggle.com/c/landmark-recognition-challenge https://www.kaggle.com/jessicali9530/lfw-dataset 卐侧 卐侧 歗⫷ID 歗⫷ID -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. ➂嵲䨌遭 • ر٦ةָ搀ֽל꧊לְְׄׯזְ • Web Crawling ח״歗⫷ ꧊הCrowdsourcing (Amazon Mechanical Turk, etc.) ח״،ظذ٦ءّٝ⡲噟 • ֲֿ׃ג ImageNet ָ钰欰 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Amazon SageMaker Ground Truth • ر٦ةחٓكٕ (Ground Truth) ➰♷ׅ،ظذ٦ءّٝ佄䴂 • ⟃♴ך4ةأؙחכذٝفٖ٦زָ欽䠐ׁגֶծ荈⡲〳腉 • ٓكٕ➰♷ׅٙ٦ؕ٦כծAmazon Mechanical Turkծ 㢩鿇كٝتծفٓ؎ك٦زזث٦يך3אַ鼅ץ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. ➂嵲䨌遭ךꣲ歲 ➂㆞䫎Ⰵ׃ג鍑寸׃זְ㜥さָ֮ 1➂ך겣钠陎ׅךח侧涰卐 乆䕦ׅךכءأذي涸ח灶笼 겣歗⫷钠陎 https://ai.googleblog.com/2018/03/google-landmarks-new-dataset-and.html ٓٝسو٦ؙ钠陎 קהו濼גְזְٓٝسو٦ؙ ך歗⫷꧊זְ https://link.springer.com/chapter/10.1007/978-3-319-25958-1_8 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. ➙傈ך鑧겗 㼭鋉垷ז歗⫷ر٦ةإحزח Deep Learning 黝欽׃ְ 黝欽דֹה⡦ֲָ׃ְ 1. ر٦ةָזֻג镘גְ➂佸ִ〳腉䚍֮կ 2. Deep Learning 黝欽ׅהծ葺ְ礵䏝ָדַ׃ זְկ 3. ⟎חر٦ة꧊ה׃גծך⸤⸂㣐ֹֻ 幾דַֹ׃זְկ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 㼭鋉垷歗⫷ر٦ةإحزח㼎ׅ،فٗ٦ث Data Augmentation ד歗⫷㟓װׅ 㼭鋉垷ז歗⫷ ח䓼ְٌرٕ ⡲ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 㼭鋉垷歗⫷ر٦ةإحزח㼎ׅ،فٗ٦ث Data Augmentation ד歗⫷㟓װׅ 㼭鋉垷ז歗⫷ ח䓼ְٌرٕ ⡲ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Data Augmentation הך歗⫷⸇䊨׃גծⵃ欽〳腉ז歗⫷㟓װׅ倯岀 ⯋歗⫷ 䊩〸鯄 ♳♴鯄 㔐鯄 [⽃秪ז⸇䊨] -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Data Augmentation GAN ⯋歗⫷ 欰䧭歗⫷ ⸇䊨٥嫰鯰粸鵤׃ծ ⯋歗⫷ה⼒ⴽָאַזְ ״ֲז歗⫷荈⹛欰䧭 嫰鯰 A. Antoniou, et al., Data Augmentation Generative Adversarial Networks, arXiv:1711.04340, 2017 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Data Augmentation GAN - Result • 垥彊涸ז⸇䊨הGANח״⸇䊨ך嫰鯰 • 歗⫷卐侧ָ㼰זְקוGANך⸬卓ָ״ֻ植גְ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. AutoAugment • 圫ղז⸇䊨倯岀ָ֮זַדծ歗⫷ח״ג黝ⴖז⸇䊨倯岀כ麩ֲכ׆ • 歗⫷ַ荈⹛ד黝ⴖז⸇䊨倯岀ⴻ倖׃ג黝欽 Ekin D. Cubuk, et al., AutoAugment: Learning Augmentation Policies from Data, arXiv:1805.09501, 2018 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. AutoAugment 然桦ⴓ䋒ח䖞ג ر٦ة欰䧭 • 㷕统 • 鐰⣣ 䓼⻉㷕统ד黝ⴖז⸇䊨倯岀㷕统،ٕ؞ٔؤيכPPO 㜠ꂹ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. AutoAugment - Result Cifer-10דך穠卓 (Error rate [%] ) Autoaugment ⵃ欽ֿׅהדerror rateָ鯪幾 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 㼭鋉垷歗⫷ر٦ةإحزח㼎ׅ،فٗ٦ث Data Augmentation ד歗⫷㟓װׅ 㼭鋉垷ז歗⫷ ח䓼ְٌرٕ ⡲ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Augmentation vs Deep Learning Model • DAGANך穠卓鋅הծ然ַחDAGANכ剣⸬ • ׃ַ׃ծDeep Learning ךٌرٕך鼅ן倯ךקֲָꅾ銲חִ DAGANד34% 礵䏝何㊣ DAGANד0.5% 礵䏝何㊣ ٌرٕד嚊י礵䏝 ָ寸תծ Augmentationד 礵䏝ָ㼰׃何㊣ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 㼰ꆀر٦ةإحزח㼎ׅDeep Learning ך倯ꆙ • Ⰻؙٓأח㼎ׅⴓ겲ٌرٕכ⡲זְ⡲זְ • 傀濼ך歗⫷ךֲ⡂גְך罋ח׃גⴓ겲ׅ (k-NNك٦أהㄎל 㢳ؙٓأⴓ겲 ٌرٕ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 稱➜׃ְⰻ㺁 • FaceNet: 겣歗⫷钠陎 F. Schroff, et al., FaceNet : A Unified Embedding for Face Recognition and Clustering, CVPR 2015 • Matching Network: 㼰侧歗⫷钠陎 (A few shot learning) O. Vinyals, et al., Matching Networks for One Shot Learning, arXiv:1606.04080 • AnoGAN: 歗⫷殯䌢嗚濼 T. Schlegl, Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery, IPMI2017 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. FaceNet F. Schroff, et al., FaceNet : A Unified Embedding for Face Recognition and Clustering, CVPR 2015 • 1➂֮ך겣歗⫷卐侧ָ㼰זְ㜥さך겣钠陎 • 겣歗⫷嫰鯰׃ג • ず♧➂暟ך겣כזץֻ鵚ֻח • ➭➂ך겣כזץֻ黅ֻח ז״ֲח瑞חꂁ縧ׅկ ず♧➂暟→鵚ְ ➭➂→黅ְ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. FaceNet, 瑞חꂁ縧הכ • ꂁ縧ׅך䏟垥ָ䗳銲 • 䏟垥鎘皾ׅ鿇ⴓאֻ 䏟垥 䏟垥鎘皾鿇ⴓ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. FaceNet, ⽃秪ז䏟垥鎘皾 • ぐؾؙإٕך⦼《⳿׃ג䏟垥ח׃ג׃תֲ • 2如⯋邌爙חׅזծ如⯋㖇簭ׅ … • 胜兝װ⯔٥䕦ך䕦갟湫䱸「ֽג׃תֲ • ず♧➂暟ך暴䗙ծ➭➂הך䊴殯ֲתֻ僥׃䏟垥ח׃ְ 铬겗 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. FaceNet, Triplet Loss ك٦أ歗⫷ծず♧➂暟ך歗⫷ծ➭➂ך歗⫷ך3אך䏟垥鎘皾 䏟垥鎘皾欽 طحزٙ٦ؙ Inception ك٦أ ず♧ ➭➂ 剑ⴱכ➭➂ךקֲָ鵚ְַ D(➭➂) D(ず♧) ず♧➂暟鵚ֻ׃ְ D(➭➂) < D(ず♧) Triplet Loss = | D(ず♧) – D(➭➂) | → min. -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. FaceNet - Result • Labeled Faces in the Wild dataset • 겣钠陎ך礵䏝ָ99.63%חⵋ麦 • Youtube Faces DB • FaceNetד95.12% • 2014䎃ךDeepFaceָ91.4% -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Matching Network ך嚊銲 • 㷕统ر٦ةך♧鿇Support setה嫰鯰׃גծ⡂גְ歗⫷ךٓ كٕ⳿⸂ׅאתMatching) • ⡂גְ䏝さְ Deep Learning ד㷕统ׅ Support set ⴓ겲㼎韋 ⴓ겲㼎韋כ Ӎ ח⡂גְ O. Vinyals, et al., Matching Networks for One Shot Learning, arXiv:1606.04080 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Matching Network ך嚊銲 侧䒭דءٝفٕח剅ֻה Support set ⴓ겲㼎韋 ⴓ겲㼎韋כ Ӎ ח⡂גְ ! " ("$ , &$ ) ! & ! & = ) $ *("$ , ! ")&$ *("$ , ! ") ٓكٕ&$חأ؝،* "$ , ! " ꅾ➰ֽ׃ג駈ׅ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Matching Network ך㷕统 Support set Batch Full dataset Support setהBatchٓٝتيד䒷ְגBatchָ姻׃ֻⴻ㹀ׁ״ֲחׅ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Matching Network ך㷕统 Support set Batch Full dataset 侧䒭ד剅ֻה ! = argmax ( )*~, )-~*,/~* [ 1 2,3 ∈/ log 7( (9|;, <)] -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. Matching Networkך穠卓 • 傀㶷ךطحزٙ٦ؙSiamese Net ה嫰ץג礵䏝ָ葺ְ • Siamese Net תծ2אךر٦ة嫰鯰׃זָ㷕统ٌׅرٕ • Fine-tuning כ䗳׆׃ֲתְַֻזַ (♴ך؛٦أדכoverfitting) -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. AnoGAN: GANⵃ欽׃歗⫷殯䌢嗚濼 • 氫䝕ך嗚⳿זוך殯䌢ر٦ةכ姻䌢ر٦ةח嫰ץה㼰זְ • 殯䌢ה姻䌢ⴓֽٌرٕ⡲ךכꨇ׃ְ • 姻䌢ر٦ةך㷕统׃גծ劢濼ךر٦ةָ姻䌢ر٦ةַ וְַֻֽꨄגְַ鋅 T. Schlegl, Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery, IPMI2017 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. 歗⫷嫰鯰ח״殯䌢嗚濼،فٗ٦ث: GAN 姻䌢禸ך ر٦ة欰䧭ׅ طحزٙ٦ؙ 姻䌢禸ך歗⫷ה 劢濼ך歗⫷嫰鯰 ׃גⴻⴽׅطحزٙ٦ؙ -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. AnoGAN - Result • 嫰鯰䩛岀㛇劤涸חכ姻䌢禸㷕统׃גծ劢濼歗⫷ה嫰 鯰ׅ䩛岀 • 姻䌢禸ך㷕统ָCAE (Conditional AutoEncoder)ծ GANָׅծ䏟垥ך״ֲזך䖤ךָ湡涸 -
© 2018, Amazon Web Services, Inc. or its Affiliates. All
rights reserved. תה • ת׆כ知⽃זAugmentationׅ • Ⰻؙٓأⴓ겲ٌׅرٕ⡲ךדכזֻծر٦ة嫰鯰 דؙٓأ寸״ֲחׅկ • 嫰鯰ךחر٦ةך瑞ꂁ縧寸ٌرٕ㷕统ׅկ • ٗأ䊨㣗ׅ倯岀 (Triplet loss) • ر٦ة嫰鯰ٌׅرٕךתת㷕统 (Matching network) • GANךDiscriminatorד嫰鯰ׅ (AnoGAN) • ⡭⸂ָ֮לAugmentation갹䓸կ