端端舝

Copyright (C) 2018 DeNA Co.,Ltd. All Rights Reserved.Copyright (C) 2018 DeNA Co.,Ltd. All Rights Reserved.
July 25, 2018
Kentaro Tachibana
AI System Dept.
DeNA Co., Ltd.
ICML2018掂心頗
Overview & GANs

Copyright (C) 2018 DeNA Co.,Ltd. All Rights Reserved.
赻撩畿賡
? 靡ゴㄩ橖翩怮檔
♂ 2008?17爛𡌂皏旃噶嶱逃本件正奈
♂ 秞汒磁傖及旃噶嶱逃卞𣶹岈
♂ 2014?17爛9堎ロ�籵陓旃噶辻�ㄗNICTㄘ堤砃
♂ 秞汒楹偈失皿伉VoiceTra及秞汒磁傖毛童絞
♂ 2017爛10堎 DeNA⻌扦
♂ 失瓦丟平乓仿汜傖卞觓歹月ㄗhttps://dena.com/intl/anime-generation/ㄘ
? 端端舝 share
♂ 汒窐劐遙及衙猁午郔陔忒楊及庄賡
♂ Icassp2018 逃桶統樓�豢 FFTNet, Tactron2畿賡

失斥尼件母
? ICML卞勾中化
? 衙猁
? �杅
? GANs憝窣�恅及�A砃 -task�e卞煦�-
? GANs憝窣井日�恅畿賡 3�
? 引午戶?垀覜

ICML2018衙猁 (1/4)
? ICML午反
? International conference on Machine Learning
? 辻迮悝�炵及玄永皿市件白央伊件旦及珨勾﹜踏隙匹35隙醴
? 掛頗荽﹜Workshop(Theoretical Foundations and Applications of Deep Generative
Models) 卞統樓仄引仄凶
? 頗� :Stockholmsmassan Stockholm, SWEDEN
? Stockholmsm?ssan, Stockholm SWEDEN
翋卅旦禾件扔奈

ICML2018衙猁 (2/4)
? き𦲀堤桯
井卅曰呏暿卅荂砓﹝
翍氪午眻諉�六月惸笭卅�匹丐勻凶﹝
NVIDIA及犯乒互俇傖僅互詢仁﹜荂砓卞紹勻凶﹝
? 旦奈由奈旦伕奈乒奈扑亦件
? 賒砉及左皮斥尼弁玄�摩

ICML2018衙猁 (3/4)
? �杅
? 芘詨杅: 2473﹜逃桶璃杅: 621 ㄗ妸亼薹: 25.1%ㄘ
? Tutorial杅: 9﹜workshop杅: 67
? 芘詨杅及芢痄
? 芘詨杅互﹜酖爛井日1.45捷卞�樓

ICML2018衙猁 (4/4)
? Session�e及芘詨杅
? 疋件示弗仄化仄引勻化引允互﹜奻弇反
1. Neural Network Architectures
2. Reinforcement Learning
3. Deep Learning (Theory)
4. Deep Learning (Adversarial)
5. Optimization (Convex)
1., 2. 互㠙仃化云曰﹜�g极及session午仄化反
�趙悝�互�中荂砓

39%
25%
18%
14%
4%
悝�失伙打伉朮丞蜊囡
Network architecture及枑偶
汜傖磐彆及啐�硌㻢及枑偶
𡛟蚚
悝�及煦昴
GANs憝窣�恅及�A砃每 task�e卞煦� -
? GANs憝窣�恅杅: 28掛
? Generative models,
Network architectures,
Network architectures (Adversarial),
workshop井日喲堤
? 黃剿匹及瓚剿及凶戶﹜穢木互丐月井手
仄木引六氏
? �砓�恅反葆嶕卞𡧳徭

畿賡�恅
1. 悝�失伙打伉丞蜊囡
? Which Training Methods for GANs do actually Converge?
2. Network architecture及枑偶
? ??RadialGAN: Leveraging multiple datasets to improve target-specific predictive models using Generative
Adversarial Networks
3. 汜傖啐�硌㻢及枑偶
? Assessing Generative Models via Precision and Recall

Which Training Methods for GANs do actually Converge?
? 掖劓
? GAN反﹜�薯卅汜傖乒犯伙分互﹜悝�互嬪褣
? 悝�及伃�互俇�卞燴賤今木凶歹仃匹反卅中
? 杅嗣仁及悝�假隅趙忒楊互枑偶今木化中月 (WGAN [1]﹜WGAN-GP [2]卅升)
? 升及忒楊互灍蕣卞�旰仄化中月及井毛捼𦷰允月
? �瓬
1. �眢卅灍歠卞方曰﹜unregularized GAN悝�反中勾手�旰反允月歹仃匹卅中仇午毛復庲
2. 郔陔及悝�假隅趙忒楊卞勾中化及�旰俶及捼𦷰
1. WGAN﹜WGAN-GP反中勾手�旰允月歹仃匹反卅中
2. Instance noise [3]﹜zero-centered gradient penalties [4] 反中勾手�旰允月
3. 扑件皿伙卅gradient penalties及枑偶
4. 灍歠腔啐�卞方曰﹜3.及�旰假隅俶毛復庲

Dirac-GAN
? Dirac-GAN及爀⻌
? 意〝及GAN悝�假隅失伙打伉朮丞毛�旰假隅俶毛復庲允月凶戶﹜�眢及醴腔憝杅毛𨃨�
? 醴腔憝杅
? GAN
? Dirac-GAN
? 淩及煦票﹜generator毛汛憝杅﹜discriminator毛�倛憝杅匹桶政
奻�毛測⻌允月午
? ? ? ￣ true data distribution
? ? ￣ generator distribution
? ? ? ￣ disctriminator
? ? ? ￣ generator
? ￣ parameters of the generator
肉￣ parameters of the discriminator
, ? ? = ,
? ? ? = ? ? = ? , ? ? ? = 0 ? ≧ ?
Vanilla-GAN:
WGAN: ? ? = ?

Dirac-GAN
? 暫湔及悝�假隅趙忒楊卞云仃月�旰杻俶及褫�趙
? Gradient vector field及爀⻌
? Jacobian及嘐衄�及灍杅窒互�化�匹丐勻凶�磁﹜ⅸ算袨颷午卅曰﹜假隅允月↙ �旰允月 [5]
Jacobian
ⅸ算袨颷ㄩ媆嶲腔
卞劐趙仄卅中袨颷
WGAN:
嘐衄� : ㊣? 午卅曰﹜假隅午反卅日卅中
Zero-centered gradient penalty:
嘐衄� : ?
?
2
㊣
?2
4
? ?∩ 0 2 ?, ? > 0 匹﹜嘐衄�反�化�午卅月凶戶﹜�旰允月
毛discriminator卞袚樓
?
?
?
?

Dirac-GAN
? 公及坻忒楊及�旰杻俶
? Non-saturating GAN﹜instance noise
匹�旰互復庲匹五月
? 仄井仄﹜non-saturating GAN反
準都卞腴厒

Regularized gradient vector field
? Discriminator及淏�t趙�午仄化﹜眕狟毛醴腔憝杅卞袚樓允月仇午毛枑偶
? Regularized gradient vector field
? 淏�t趙�卞方曰﹜假隅卞�旰允月

灍歠腔啐�及磐彆
? Imagenet

灍歠腔啐�及磐彆
? CelebA-HQ

Tips
? Generator午Discriminator反﹜肮媆卞由仿丟奈正載陔六內﹜⑵�內勾載陔允月
? momentum反妏丹卅
? 悝�假隅趙及凶戶卞反﹜淏�t趙毛妏丹
? Discriminator卞�允月zero-centered gradient penalties反謎中磐彆毛爀仁
↙ 翍氪及github page: https://github.com/LMescheder/GAN_stability

??RadialGAN: Leveraging multiple datasets to improve target-specific
predictive models using Generative Adversarial Networks
? �觳偞隅
? 遞氪ロ�井日杻隅及瓷𠸎及逃汜薹﹜侚厗薹毛軑䛐仄凶中
? 瓷埏仍午卞偞�手�卅月凶戶﹜遞氪ロ�卞⑴曰互堤月
? 掝觳
? Feature mismatch
? 悝�犯奈正本永玄午醴㻢犯奈正本永玄匹杻釾講互�卅月
? Distribution mismatch
? 悝�犯奈正本永玄午醴㻢犯奈正本永玄匹﹜肮元杻釾講匹手復薹煦票互�卅月
遞氪D
爛溵
极笭
網柲厒僅
伊件玄必件
MRI
遞氪C
爛溵
极笭
網柲厒僅
伊件玄必件
MRI
遞氪B
爛溵
极笭
網柲厒僅
伊件玄必件
MRI
遞氪A
爛溵
极笭
網柲厒僅
伊件玄必件
MRI
遞氪C
爛溵
极笭
遞氪B
爛溵
极笭
遞氪A
爛溵
极笭
遞氪C
爛溵
极笭
遞氪B
爛溵
极笭
悛㇌
遞氪A
爛溵
极笭
悛㇌
侚厗薹
骨逃痌薹
ˋ
侚厗薹
骨逃痌薹

? 失奶犯失
? CycleGAN矛奈旦卞恚杅GAN毛瞳蚚
? 遞氪犯奈正毛サ婓諾嶲卞mapping仄﹜僕籵諾嶲匹㜳丹
? 猁匼
? Encoder ??: ?(?)
℅ ? ↙ ?,
? Decoder
? Discriminator
? Loss
1. Adversarial loss
2. Cycle-consistency loss
3. Optimization
??RadialGAN 失伙打伉朮丞
Di :瓷埏
X(i) :遞氪犯奈正
Y(i) :軑䛐仄凶中犯奈正
Z :サ婓諾嶲
?? = ?? ?(?), ? ﹋ ?
??: ? ↙ ?(?) ℅ ?, ?(?)
, ? = ?? ?? ﹋ ?(?)
℅ ?
? ?(?)
, ? = ? ?(?)
, ? ﹋ ????: ?(?)
℅ ? ↙ [0, 1],

灍歠腔啐�
? 灍歠沭璃
? 灍歠犯奈正 14瓷埏匹528�井日13279�及遞氪犯奈正
? 悝�犯奈正 500�及遞氪犯奈正﹜氾旦玄犯奈正公及坻紹曰及犯奈正
? 磁�216意及杻釾講ㄗ跪瓷埏匹�卅月ㄘ匹﹜堤薯ㄩ1爛摽及侚厗薹毛軑䛐
? 灍歠磐彆
? GAN及汜傖磐彆毛伕斥旦永氾奴弁隙䔝仄化﹜啐�
AUC: ?Area Under ROC Curve
APR: ?Average Precision Recall
Baseline
Gain

�瓷埏匹及灍歠磐彆
RadialGAN反幻午氏升及label匹AUC﹜APR互�樓仄化中月

Assessing Generative Models via Precision and Recall
? 掖劓
? 汜傖乒犯伙及磐彆毛隅講腔卞啐�允月仇午反褣仄中
? 蚧僅及方丹卅躇僅芢隅匹反﹜蚧僅互狟互勻凶井日午中勻化﹜ⅲ斮?嗣�俶互蜊囡仄化月歹仃
匹反卅中 [6]
? 郔陔及啐�忒楊 Inception score (IS) [7]午 ?Frechet inception distance (FID) [8] 匹反﹜�卅月嵞匹及
啐�互堤懂卅中
? �瓬
? 復薹煦票卞�允月precision (ⅲ斮)午recall (嗣�俶)及隅膽毛枑偶
? GAN及mode dropping﹜model inventing毛隅講腔卞䛐隅
? 暫湔及汜傖乒犯伙毛枑偶楊及嵞匹掀廌

肮�及FID及賒砉掀廌
? 升切日互謎中及井ˋ

肮�及FID及賒砉掀廌
ⅲ斮 ?﹜嗣�俶 ℅ ⅲ斮 ℅﹜嗣�俶 ?
ⅲ斮午嗣�俶毛煦褩匹五卅中

Precision and Recall for Distribution (PRD)
? Precision (ⅲ斮)
? 悝�乒犯伙及煦票Q互升及最僅﹜淩及煦票 P 毛汜傖堤懂化中月井ˋ
? Recall (嗣�俶)
? 淩及煦票 P 互升及最僅﹜悝�乒犯伙及煦票 Q 毛汜傖堤懂化中月井ˋ
P : 淩及煦票
Q : 悝�乒犯伙及煦票
High precision
Low precision
Low recall High recall

Precision and Recall for Distribution (PRD)
? 煦票 P 午 Q 反眕狟及�卞桶六月
? PRD(Q, P)反汐, 汕毛蚚中化眕狟匹隅膽今木月
? �g�趙仄﹜褒釱㻢匹桶�允月午
? ? ? ?
?
P : 淩及煦票
Q : 悝�乒犯伙及煦票
Recall(汕)卞憝允月
loss
Precision(汐)卞憝允月
loss
? = ?? + (1 ? ?)? ?
? = ?? + (1 ? ?)? ?
??? ?, ? = { ?? ? , ?? ? | ? ﹋ 0, ﹢ , ? ﹋ 0, 1 }
? ? =
?﹋次
??? ?? ? , ? ? ,
偩隴反appendix A.2統桽
汐 :precision﹜汕 :recall
? ? =
?﹋次
??? ? ? ,
? ?
?
??? ?, ? = { ? ? , ? ? | ? ﹋托}, 托 = tan
?
? + 1
?
?
2
| ? = 1, 2, ? ? , ? ﹋ ?
次反衄癹袨颷諾嶲

灍歠腔啐� with MNIST and CelebA
ⅲ斮 ?﹜嗣�俶 ℅ ⅲ斮 ℅﹜嗣�俶 ?
Precision 酘 > 衵﹜recall 酘 < 衵午卅曰﹜ⅲ斮午嗣�俶毛啐�匹五化中月

Mode dropping and inventing及灍歠腔啐�
CIFAIR-10毛蚚中化﹜淩及煦票P及弁仿旦杅毛5卞嘐隅仄凶
悝�煦票 Q及弁仿旦杅毛劐尹化﹜mode dropping and inventing匹及伃�毛掀廌
IS反�g捼�樓
FID反mode dropping﹜mode inventing卞鏗覜
PRD反弁仿旦杅1-4分午p互詢﹜r互腴﹜弁仿旦杅6-10反公及欄

汜傖乒犯伙及灍歠腔啐� with Fashion-MNIST
? 800及汜傖乒犯伙ㄗ7GANs午VAE x 100parameters) 及 F汕旦戊失毛啐�
汕卞方曰﹜precision (p)午recall (r)毛笭心葆仃

引午戶?垀覜
? 引午戶
? ICML2018卞勾中化﹜衙猁午�杅卞勾中化畿賡
? GANs憝窣�恅及�A砃午﹜3�毛畿賡
? 垀覜
? GANs及煦珧匹反﹜悝�失伙打伉朮丞及蜊囡﹜𡛟蚚卅升田仿件旦方仁旃噶今木化中月﹝
? 撮胍筳桯及旦疋奈玉反午化手厒中﹜袚中化俴井木卅中�﹜失件氾瓜毛�勻化中仁
? US及湮忒ITき𦲀互湮悝午僕翍匹逃桶仄化中凶�恅互準都卞嗣井勻凶
? 桯尨頗匹手搪婡腔卞伉弁伙奈玄魂�毛仄化中月荂砓

統蕉恅瓬
[1] ?Arjovsky, M. and Bottou, L. Towards principled methods for training generative adversarial networks. CoRR, abs/1701.04862, 2017.
[2] ?Gulrajani, I., Ahmed, F., Arjovsky, M., Dumoulin, V., and Courville, A. C. Improved training ofwasserstein gans. In Advances in Neural
Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, 4-9 December 2017, Long Beach,
CA, USA, pp. 5769每5779, 2017.
[3] ?S?nderby, C. K., Caballero, J., Theis, L., Shi, W., and Husz芍r, F. Amortised MAP inference for image super- resolution. CoRR,
abs/1610.04490, 2016.
[4] ?Roth, K., Lucchi, A., Nowozin, S., and Hofmann, T. Stabi- lizing training of generative adversarial networks through ?regularization. In
Advances in Neural Information Pro- cessing Systems 30: Annual Conference on Neural In- formation Processing Systems 2017, 4-9
December 2017, Long Beach, CA, USA, pp. 2015每2025, 2017.
[5] ?Mescheder, L. M., Nowozin, S., and Geiger, A. The numer- ics of gans. In Advances in Neural Information Process- ing Systems 30: Annual
Conference on Neural Informa- tion Processing Systems 2017, 4-9 December 2017, Long Beach, CA, USA, pp. 1823每1833, 2017.
[6] ?Salimans, T., Goodfellow, I., Zaremba, W., Cheung, V., Rad- ford, A., and Chen, X. Improved Techniques for Training GANs. In Advances in
Neural Information Processing Systems (NIPS), 2016.
[7] ?Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Klambauer, G., and Hochreiter, S. Gans trained by a two time-scale update rule
converge to a nash equilib- rium. In Advances in Neural Information Processing Systems (NIPS), 2017.
[8] ?Theis, L., Oord, A. v. d., and Bethge, M. A note on the evaluation of generative models. In International Conference on Learning
Representations (ICLR), 2016.

葆嶕 (1/3)
正旦弁靡 �恅
悝�失伙打伉朮丞蜊囡
mode collapse隙旌﹜悝�假隅俶砃奻﹜載陔宒
蜊囡卅升
Which Training Methods for GANs do actually Converge?
Chi-square Generative Adversarial Network
Tempered Adversarial Networks
Improved Training of Generative Adversarial Networks Using Representative
Features
A Two-Step Computation of the Exact GAN Wasserstein Distance
First Order Generative Adversarial Networks
Mixed batches and symmetric discriminators for GAN training
Stochastic Variance Reduced Gradient Optimization of Generative Adversarial
Networks
Bayesian Modelling and Monte Carlo Inference for GAN
Approximability of Discriminators Implies Diversity in GANs
On catastrophic forgetting and mode collapse in Generative Adversarial
Networks

葆嶕 (2/3)
正旦弁靡 �恅
Network architecture及枑偶
RadialGAN: Leveraging multiple datasets to improve target-specific
predictive models using Generative Adversarial Networks
GAIN: Missing Data Imputation using Generative Adversarial Nets
MAGAN: Aligning Biological Manifolds
JointGAN: Multi-Domain Joint Distribution Learning with Generative
Adversarial Nets
Augmented CycleGAN: Learning Many-to-Many Mappings from Unpaired
Data
BAGAN: Data Augmentation with Balancing GAN
Generating Multi-Categorical Samples with Generative Adversarial Networks

葆嶕 (3/3)
正旦弁靡 �恅
汜傖磐彆及啐�硌㻢及枑偶
GAN及汜傖磐彆毛諦紝啐�允月忒楊毛枑偶
A Classification-Based Study of Covariate Shift in GAN Distributions
Geometry Score: A Method For Comparing Generative Adversarial Networks
On Accurate Evaluation of GANs for Language Generation
Assessing Generative Models via Precision and Recall
Towards GAN Benchmarks Which Require Generalization
𡛟蚚
暫湔architecture毛陔凶卅煦珧尺羥蚚
NetGAN: Generating Graphs via Random Walks
BinGAN: Learning Compact Binary Descriptors with a Regularized GAN
MolGAN: An implicit generative model for small molecular graphs
Generative Buyer-Seller Adversarial Networks
悝�及煦昴
悝�假隅俶卞砃仃凶煦昴
Nonstationary GANs: Analysis as Nonautonomous Dynamical Systems

端端舝

梆釵鳥梭2018掂心頗喳棗措梗娶措勳梗滄&硃鳥梯;勞插捧莽

Recommended

More Related Content

What's hot (20)

Similar to 梆釵鳥梭2018掂心頗喳棗措梗娶措勳梗滄&硃鳥梯;勞插捧莽 (20)

More from Kentaro Tachibana (6)

Recently uploaded (7)

梆釵鳥梭2018掂心頗喳棗措梗娶措勳梗滄&硃鳥梯;勞插捧莽

Editor's Notes