label-noise robust generative adversarial networks · label-noise robust generative adversarial...

Takuhiro Kaneko1 Yoshitaka Ushiku1 Tatsuya Harada1, 2

1The University of Tokyo 2RIKEN

Noisy labeled Conditioned on clean labels

Training data rcGAN

Code

Label-Noise Robust Generative Adversarial Networks

Talk

Objective: Label-noise robust conditional image generation

2

Our goal is to construct a label-noise robust conditional image generatorthat can reproduce clean labeled data (a)even when noisy labeled data (b) are only available during the training.

(a) Clean labeled data (b) Noisy labeled data

AvailableUnavailable

Challenge: Limitation of naïve conditional generative models

3

Naïve conditional generative models (e.g., AC-GAN [1], cGAN [2, 3])attempt to construct a generator conditioned on observable labels(i.e., noisy labels in this case).

Noisy labeled Conditioned on noisy labels

Training data cGAN

[1] Odena et al. ICML 2017. [2] Mirza & Osindero. arXiv 2014. [3] Miyato & Koyama. ICLR 2018.

AC-GAN: auxiliary classifier GAN, cGAN: conditional GAN

Contribution: Proposal of label-noise robust GANs

4

To overcome this limitation, we propose label-noise robust GANs (rGANs) thatcan construct a generator conditioned on clean labelseven when trained with noisy labeled data.

Conditioned on clean labels

rcGAN

Noisy labeled

Training data

rcGAN: label-noise robust conditional GAN

We incorporate a noise transition model into naïve conditional GANs [1, 2, 3].

Two variants

DDD/CD/C

Main idea: Incorporation of noise transition model

5

p(y|y)<latexit sha1_base64="U76tA9Mz7kpaJIgkixOqENDvr3w=">AAACDnicbZDJSgNBEIZr4hbjFhW8eGkShIgQZrzoMejFYwSzQCaEnk4nadKz0F0jDJO8g48gXvXsTbz6Ch59EzvLwST+UPDxVxVV/F4khUbb/rYya+sbm1vZ7dzO7t7+Qf7wqK7DWDFeY6EMVdOjmksR8BoKlLwZKU59T/KGN7yd9BuPXGkRBg+YRLzt034geoJRNFYnfxKVXBSyy9NkTEbEHVA0dN7JF+2yPRVZBWcOxUrBvXgGgGon/+N2Qxb7PEAmqdYtx46wnVKFgkk+zrmx5hFlQ9rnLYMB9blup9P/x+TMOF3SC5WpAMnU/buRUl/rxPfMpE9xoJd7E/O/XivG3nU7FUEUIw/Y7FAvlgRDMgmDdIXiDGVigDIlzK+EDaiiDE1kC1c8f2wycZYTWIX6ZdkxfG/CuYGZsnAKBSiBA1dQgTuoQg0YjOAFXuHNerLerQ/rczaaseY7x7Ag6+sXDAKeCA==</latexit><latexit sha1_base64="bljVtG6foLbmWotaiB3KIQEccSo=">AAACDnicbZC7SgNBFIZn4y3G26pgYzMkCBEh7NpoGbSxjGASIbuE2dlJMmT2wsxZYdnsO1j4BLZa24mtr5DSN3FyKUziDwc+/nMO5/B7seAKLGtsFNbWNza3itulnd29/QPz8KilokRS1qSRiOSjRxQTPGRN4CDYYywZCTzB2t7wdtJvPzGpeBQ+QBozNyD9kPc4JaCtrnkSVx3gwmdZmuMRdgYENJ13zYpVs6bCq2DPoVIvOxcv43ra6Jo/jh/RJGAhUEGU6thWDG5GJHAqWF5yEsViQoekzzoaQxIw5WbT/3N8ph0f9yKpKwQ8df9uZCRQKg08PRkQGKjl3sT8r9dJoHftZjyME2AhnR3qJQJDhCdhYJ9LRkGkGgiVXP+K6YBIQkFHtnDFC3Kdib2cwCq0Lmu25nsdzg2aqYhOURlVkY2uUB3doQZqIopG6BW9oXfj2fgwPo2v2WjBmO8cowUZ378sRp+O</latexit><latexit sha1_base64="rdvUqbcmddb3mFWLiBIwQdy5QAc=">AAACDnicbZDLSsNAFIYnXmu9RQU3bgaLUDclcaPLohuXFewFmlAmk0k7dHJh5kQIad7BZ3Cra3fi1ldw6Zs4bbOwrT8c+PjPOZzD7yWCK7Csb2NtfWNza7uyU93d2z84NI+OOypOJWVtGotY9jyimOARawMHwXqJZCT0BOt647tpv/vEpOJx9AhZwtyQDCMecEpAWwPzNKk7wIXP8qzAE+yMCGi6HJg1q2HNhFfBLqGGSrUG5o/jxzQNWQRUEKX6tpWAmxMJnApWVJ1UsYTQMRmyvsaIhEy5+ez/Al9ox8dBLHVFgGfu342chEploacnQwIjtdybmv/1+ikEN27OoyQFFtH5oSAVGGI8DQP7XDIKItNAqOT6V0xHRBIKOrKFK15Y6Ezs5QRWoXPVsDU/WLXmbZlOBZ2hc1RHNrpGTXSPWqiNKJqgF/SK3oxn4934MD7no2tGuXOCFmR8/QLibZx/</latexit>

Noise transition model

Clean labelNoisy label

AC-GAN [1] rAC-GAN cGAN [2, 3] rcGAN


D/C

Auxiliary classifier GAN [1]

Generator:

Discriminator/Classifier: /

G

Baseline: AC-GAN

6

G(z, yg)<latexit sha1_base64="R7RPelWiq6Kk4lORxCVK9XU7AP4=">AAACCXicbZC7SgNBFIbPxluMt42WNkOCEFHCro2WQQstI5gLJDHMTibJkJndZWZWWZd9Ap9BsNLaTmx9CkvfxEliYRJ/OPDxn3M4h98LOVPacb6szNLyyupadj23sbm1vWPnd+sqiCShNRLwQDY9rChnPq1ppjlthpJi4XHa8EYX437jjkrFAv9GxyHtCDzwWZ8RrI3VtfOXpaTtCfSQHqMkTm8Hh1276JSdidAiuL9QrBTaR88AUO3a3+1eQCJBfU04VqrlOqHuJFhqRjhNc+1I0RCTER7QlkEfC6o6yeT1FB0Yp4f6gTTlazRx/24kWCgVC89MCqyHar43Nv/rtSLdP+skzA8jTX0yPdSPONIBGueAekxSonlsABPJzK+IDLHERJu0Zq54IjWZuPMJLEL9pOwavjbhnMNUWdiHApTAhVOowBVUoQYE7uEJXuDVerTerHfrYzqasX539mBG1ucPIHabRw==</latexit><latexit sha1_base64="oafkp0GnlyKHqqpdloXlNUofj28=">AAACCXicbZC7TsMwFIadcivllsLIYrVCKgJVCQuMFQwwFolepCZUjuu0Vu0ksh1QiPIErKysMLMhVp6iI2+Cexloyy8d6dN/ztE5+r2IUaksa2TkVlbX1jfym4Wt7Z3dPbO435RhLDBp4JCFou0hSRgNSENRxUg7EgRxj5GWN7wa91sPREgaBncqiYjLUT+gPsVIaatrFq8rqeNx+JSdwjTJ7vvHXbNsVa2J4DLYMyjXSs7Jy6iW1Lvmj9MLccxJoDBDUnZsK1JuioSimJGs4MSSRAgPUZ90NAaIE+mmk9czeKSdHvRDoStQcOL+3UgRlzLhnp7kSA3kYm9s/tfrxMq/cFMaRLEiAZ4e8mMGVQjHOcAeFQQrlmhAWFD9K8QDJBBWOq25Kx7PdCb2YgLL0Dyr2ppvdTiXYKo8OAQlUAE2OAc1cAPqoAEweASv4A28G8/Gh/FpfE1Hc8Zs5wDMyfj+BUC6nM0=</latexit><latexit sha1_base64="e/EvxuwTJ3maU0EAsLZWtKGhEHg=">AAACCXicbZDLSsNAFIYn9VbrLdWlm8EiVJCSuNFl0YUuK9gLtLFMppN26MwkzEyUGPIEPoNbXbsTtz6FS9/EaZuFbf3hwMd/zuEcfj9iVGnH+bYKK6tr6xvFzdLW9s7unl3eb6kwlpg0cchC2fGRIowK0tRUM9KJJEHcZ6Ttj68m/fYDkYqG4k4nEfE4GgoaUIy0sfp2+bqa9nwOn7JTmCbZ/fCkb1ecmjMVXAY3hwrI1ejbP71BiGNOhMYMKdV1nUh7KZKaYkayUi9WJEJ4jIaka1AgTpSXTl/P4LFxBjAIpSmh4dT9u5EirlTCfTPJkR6pxd7E/K/XjXVw4aVURLEmAs8OBTGDOoSTHOCASoI1SwwgLKn5FeIRkghrk9bcFZ9nJhN3MYFlaJ3VXMO3TqV+madTBIfgCFSBC85BHdyABmgCDB7BC3gFb9az9W59WJ+z0YKV7xyAOVlfv/bhmb4=</latexit>

D(x)<latexit sha1_base64="Gv+HfqGulDCUn+wu8alHXb1LATY=">AAACAHicbZDLSgMxFIbP1Futt6pLN6FFqAhlxo0ui7pwWcFesFNKJk3b0CQzJBmxDLPxGdwqLt2JW9/EpW9ielnY1h8CH/85h3PyBxFn2rjut5NZWV1b38hu5ra2d3b38vsHdR3GitAaCXmomgHWlDNJa4YZTpuRolgEnDaC4dW43nigSrNQ3plRRNsC9yXrMYKNte6vS4kfCPSYnnTyRbfsToSWwZtBsVLwT98AoNrJ//jdkMSCSkM41rrluZFpJ1gZRjhNc36saYTJEPdpy6LEgup2Mrk4RcfW6aJeqOyTBk3cvxMJFlqPRGA7BTYDvVgbm//VWrHpXbQTJqPYUEmmi3oxRyZE4++jLlOUGD6ygIli9lZEBlhhYmxIc1sCkdpMvMUElqF+VvYs39pwLmGqLBxBAUrgwTlU4AaqUAMCEp7hBV6dJ+fd+XA+p60ZZzZzCHNyvn4BiVqYSQ==</latexit><latexit sha1_base64="cbB1k1ygBI0YUzWa1QEz/S/72Vo=">AAACAHicbZDLTgIxFIbP4A3xhrp000BMMCZkxo0uibpwiYmAkSGkUwo0tJ1J2zGSyWx8Bra6dmfc+iYsfRPLZSHgnzT58p9zck7/IOJMG9cdO5m19Y3Nrex2bmd3b/8gf3hU12GsCK2RkIfqMcCaciZpzTDD6WOkKBYBp41gcDOpN56p0iyUD2YY0ZbAPcm6jGBjrafbUuIHAr2kZ+180S27U6FV8OZQrBT889G4Mqy28z9+JySxoNIQjrVuem5kWglWhhFO05wfaxphMsA92rQosaC6lUwvTtGpdTqoGyr7pEFT9+9EgoXWQxHYToFNXy/XJuZ/tWZsulethMkoNlSS2aJuzJEJ0eT7qMMUJYYPLWCimL0VkT5WmBgb0sKWQKQ2E285gVWoX5Q9y/c2nGuYKQsnUIASeHAJFbiDKtSAgIQRvMG78+p8OJ/O16w148xnjmFBzvcvqZ6Zzw==</latexit><latexit sha1_base64="2cD8/U9oFx+mLgdgp+4wOPUpte4=">AAACAHicbZC7TsMwFIZPuJZyKzCyWFRIZakSFhgrYGAsEr2INqoc12mt2k5kO4gqysIzsMLMhlh5E0beBLfNQFt+ydKn/5yjc/wHMWfauO63s7K6tr6xWdgqbu/s7u2XDg6bOkoUoQ0S8Ui1A6wpZ5I2DDOctmNFsQg4bQWj60m99UiVZpG8N+OY+gIPJAsZwcZaDzeVtBsI9JSd9Uplt+pOhZbBy6EMueq90k+3H5FEUGkIx1p3PDc2foqVYYTTrNhNNI0xGeEB7ViUWFDtp9OLM3RqnT4KI2WfNGjq/p1IsdB6LALbKbAZ6sXaxPyv1klMeOmnTMaJoZLMFoUJRyZCk++jPlOUGD62gIli9lZEhlhhYmxIc1sCkdlMvMUElqF5XvUs37nl2lWeTgGO4QQq4MEF1OAW6tAAAhJe4BXenGfn3flwPmetK04+cwRzcr5+AV/UlsA=</latexit>

C(y|x)<latexit sha1_base64="HwdQIIFTutxG8mfrVQATm5P1dgY=">AAACBHicbZC7SgNBFIbPeo3xFrW0GRKEiBB2bbQMprGMYC6QXcPsZDYZMrO7zMyKYd3WZ7BUazux9T0sfRMnl8Ik/nDg4z/ncA6/H3OmtG1/Wyura+sbm7mt/PbO7t5+4eCwqaJEEtogEY9k28eKchbShmaa03YsKRY+py1/WBv3W/dUKhaFt3oUU0/gfsgCRrA21l2tPEKPKHV9gR6y026hZFfsidAyODMoVYvu2QsA1LuFH7cXkUTQUBOOleo4dqy9FEvNCKdZ3k0UjTEZ4j7tGAyxoMpLJ19n6MQ4PRRE0lSo0cT9u5FiodRI+GZSYD1Qi72x+V+vk+jg0ktZGCeahmR6KEg40hEaR4B6TFKi+cgAJpKZXxEZYImJNkHNXfFFZjJxFhNYhuZ5xTF8Y8K5gqlycAxFKIMDF1CFa6hDAwhIeIZXeLOerHfrw/qcjq5Ys50jmJP19QsP9Zml</latexit><latexit sha1_base64="Ij2ZJYOmd8TCSllKKKQ/HEZl91I=">AAACBHicbZC7TsMwFIadcivlVmBksVohFSFVCQuMFV0Yi0QvUhMqx3Vaq7YT2Q4iCll5AgZWmNkQK+/RkTfBvQy05ZeO9Ok/5+gc/X7EqNK2PbZya+sbm1v57cLO7t7+QfHwqKXCWGLSxCELZcdHijAqSFNTzUgnkgRxn5G2P6pP+u0HIhUNxZ1OIuJxNBA0oBhpY93XKwl8gqnrc/iYnfWKZbtqTwVXwZlDuVZyz1/GtaTRK/64/RDHnAiNGVKq69iR9lIkNcWMZAU3ViRCeIQGpGtQIE6Ul06/zuCpcfowCKUpoeHU/buRIq5Uwn0zyZEequXexPyv1411cOWlVESxJgLPDgUxgzqEkwhgn0qCNUsMICyp+RXiIZIIaxPUwhWfZyYTZzmBVWhdVB3DtyacazBTHpyAEqgAB1yCGrgBDdAEGEjwCt7Au/VsfVif1tdsNGfNd47BgqzvXzA5mys=</latexit><latexit sha1_base64="6sExgRCM1tmJIywXswSrlcJrs/s=">AAACBHicbZA9T8MwEIYv5auUrwIji0WFVJYqYYGxogtjkeiH1IbKcZ3Wqp1EtoOIQlZ+AyvMbIiV/8HIP8FtM9CWVzrp0Xt3utPrRZwpbdvfVmFtfWNzq7hd2tnd2z8oHx61VRhLQlsk5KHselhRzgLa0kxz2o0kxcLjtONNGtN+54FKxcLgTicRdQUeBcxnBGtj3TeqCXpCad8T6DE7H5Qrds2eCa2Ck0MFcjUH5Z/+MCSxoIEmHCvVc+xIuymWmhFOs1I/VjTCZIJHtGcwwIIqN519naEz4wyRH0pTgUYz9+9GioVSifDMpMB6rJZ7U/O/Xi/W/pWbsiCKNQ3I/JAfc6RDNI0ADZmkRPPEACaSmV8RGWOJiTZBLVzxRGYycZYTWIX2Rc0xfGtX6td5OkU4gVOoggOXUIcbaEILCEh4gVd4s56td+vD+pyPFqx85xgWZH39AuZgmBw=</latexit>

Dataset

[1] Odena et al. ICML 2017.

D/C

Auxiliary classifier GAN [1]

Generator:


Limitation

C can fit noisy labels when trained with noisy labeled data.

G

Baseline: AC-GAN

7



C(y|x)<latexit sha1_base64="HwdQIIFTutxG8mfrVQATm5P1dgY=">AAACBHicbZC7SgNBFIbPeo3xFrW0GRKEiBB2bbQMprGMYC6QXcPsZDYZMrO7zMyKYd3WZ7BUazux9T0sfRMnl8Ik/nDg4z/ncA6/H3OmtG1/Wyura+sbm7mt/PbO7t5+4eCwqaJEEtogEY9k28eKchbShmaa03YsKRY+py1/WBv3W/dUKhaFt3oUU0/gfsgCRrA21l2tPEKPKHV9gR6y026hZFfsidAyODMoVYvu2QsA1LuFH7cXkUTQUBOOleo4dqy9FEvNCKdZ3k0UjTEZ4j7tGAyxoMpLJ19n6MQ4PRRE0lSo0cT9u5FiodRI+GZSYD1Qi72x+V+vk+jg0ktZGCeahmR6KEg40hEaR4B6TFKi+cgAJpKZXxEZYImJNkHNXfFFZjJxFhNYhuZ5xTF8Y8K5gqlycAxFKIMDF1CFa6hDAwhIeIZXeLOerHfrw/qcjq5Ys50jmJP19QsP9Zml</latexit><latexit sha1_base64="Ij2ZJYOmd8TCSllKKKQ/HEZl91I=">AAACBHicbZC7TsMwFIadcivlVmBksVohFSFVCQuMFV0Yi0QvUhMqx3Vaq7YT2Q4iCll5AgZWmNkQK+/RkTfBvQy05ZeO9Ok/5+gc/X7EqNK2PbZya+sbm1v57cLO7t7+QfHwqKXCWGLSxCELZcdHijAqSFNTzUgnkgRxn5G2P6pP+u0HIhUNxZ1OIuJxNBA0oBhpY93XKwl8gqnrc/iYnfWKZbtqTwVXwZlDuVZyz1/GtaTRK/64/RDHnAiNGVKq69iR9lIkNcWMZAU3ViRCeIQGpGtQIE6Ul06/zuCpcfowCKUpoeHU/buRIq5Uwn0zyZEequXexPyv1411cOWlVESxJgLPDgUxgzqEkwhgn0qCNUsMICyp+RXiIZIIaxPUwhWfZyYTZzmBVWhdVB3DtyacazBTHpyAEqgAB1yCGrgBDdAEGEjwCt7Au/VsfVif1tdsNGfNd47BgqzvXzA5mys=</latexit><latexit sha1_base64="6sExgRCM1tmJIywXswSrlcJrs/s=">AAACBHicbZA9T8MwEIYv5auUrwIji0WFVJYqYYGxogtjkeiH1IbKcZ3Wqp1EtoOIQlZ+AyvMbIiV/8HIP8FtM9CWVzrp0Xt3utPrRZwpbdvfVmFtfWNzq7hd2tnd2z8oHx61VRhLQlsk5KHselhRzgLa0kxz2o0kxcLjtONNGtN+54FKxcLgTicRdQUeBcxnBGtj3TeqCXpCad8T6DE7H5Qrds2eCa2Ck0MFcjUH5Z/+MCSxoIEmHCvVc+xIuymWmhFOs1I/VjTCZIJHtGcwwIIqN519naEz4wyRH0pTgUYz9+9GioVSifDMpMB6rJZ7U/O/Xi/W/pWbsiCKNQ3I/JAfc6RDNI0ADZmkRPPEACaSmV8RGWOJiTZBLVzxRGYycZYTWIX2Rc0xfGtX6td5OkU4gVOoggOXUIcbaEILCEh4gVd4s56td+vD+pyPFqx85xgWZH39AuZgmBw=</latexit>

Noisy label classifier

Noisy

Dataset

[1] Odena et al. ICML 2017.

Label-noise robust auxiliary classifier GAN

Generator:


Solution

We correct the Cʼs prediction using the noise transition model (forward correction [4]).

G

D/C

Proposal: rAC-GAN

8

Clean label classifier

Clean

Dataset

G(z, yg)<latexit sha1_base64="SOIwORCvydCa/EjDTfWAMeO4m6Y=">AAACDXicbZC7SgNBFIbPxluMt1WxshkShIgSdm20DFpoGcFcIBvD7GSSDJnZXWZmhbjsM/gK2mptJ7Y+g6Vv4uRSmMQfDnz85xzO4fcjzpR2nG8rs7S8srqWXc9tbG5t79i7ezUVxpLQKgl5KBs+VpSzgFY105w2Ikmx8Dmt+4OrUb/+QKViYXCnhxFtCdwLWJcRrI3Vtg+ui4nnC/SYniKvj3UyTO97x2274JScsdAiuFMolPPeyTMAVNr2j9cJSSxooAnHSjVdJ9KtBEvNCKdpzosVjTAZ4B5tGgywoKqVjN9P0ZFxOqgbSlOBRmP370aChVJD4ZtJgXVfzfdG5n+9Zqy7F62EBVGsaUAmh7oxRzpEoyxQh0lKNB8awEQy8ysifSwx0SaxmSu+SE0m7nwCi1A7K7mGb004lzBRFg4hD0Vw4RzKcAMVqAKBBF7gFd6sJ+vd+rA+J6MZa7qzDzOyvn4BUw2dCA==</latexit><latexit sha1_base64="Xk5dcECSM836Sq211VVs1O96jYM=">AAACDXicbZC7SgNBFIZn4y3G26pY2QwJQkQJuzZaBi20jGAukF3D7GSSDJmZXWZmhXXZZ7DxBWy1thNbnyGlb+IksTCJPxz4+M85nMMfRIwq7TgjK7e0vLK6ll8vbGxube/Yu3sNFcYSkzoOWShbAVKEUUHqmmpGWpEkiAeMNIPh1bjffCBS0VDc6SQiPkd9QXsUI22sjn1wXU69gMPH7BR6A6TTJLvvH3fsklNxJoKL4P5CqVr0Tp5H1aTWsb+9bohjToTGDCnVdp1I+ymSmmJGsoIXKxIhPER90jYoECfKTyfvZ/DIOF3YC6UpoeHE/buRIq5UwgMzyZEeqPne2Pyv145178JPqYhiTQSeHurFDOoQjrOAXSoJ1iwxgLCk5leIB0girE1iM1cCnplM3PkEFqFxVnEN35pwLsFUeXAIiqAMXHAOquAG1EAdYJCCF/AK3qwn6936sD6noznrd2cfzMj6+gFzUZ6O</latexit><latexit sha1_base64="fziiGSVr2MA5GrP28d7XcL5PqqA=">AAACDXicbZDLSsNAFIYn9VbrLSqu3AwWoYKUxI0uiy50WcFeoIllMp20Q2cmYWYixJBn8Bnc6tqduPUZXPomTtssbPWHAx//OYdz+IOYUaUd58sqLS2vrK6V1ysbm1vbO/buXltFicSkhSMWyW6AFGFUkJammpFuLAniASOdYHw16XceiFQ0Enc6jYnP0VDQkGKkjdW3D65rmRdw+JifQm+EdJbm98OTvl116s5U8C+4BVRBoWbf/vYGEU44ERozpFTPdWLtZ0hqihnJK16iSIzwGA1Jz6BAnCg/m76fw2PjDGAYSVNCw6n7eyNDXKmUB2aSIz1Si72J+V+vl+jwws+oiBNNBJ4dChMGdQQnWcABlQRrlhpAWFLzK8QjJBHWJrG5KwHPTSbuYgJ/oX1Wdw3fOtXGZZFOGRyCI1ADLjgHDXADmqAFMMjAM3gBr9aT9Wa9Wx+z0ZJV7OyDOVmfPymHm38=</latexit>


Xp(y|y)C(y|x)

<latexit sha1_base64="6w3/LMdF/zpXVPCbkPFma8JRnig=">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</latexit><latexit sha1_base64="o5Y2Vgi3HSs6u2+PicuY6utJIpY=">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</latexit><latexit sha1_base64="PPuK3/pzbUwKLJ/NqVMzlDD8zgI=">AAACLnicbZDLSsNAFIYn9VbrLerSzWAR2k1JXKjLYjcuK9gLNKVMptN26EwSZk7EEvMmPoTP4FbXggsRdz6G0zYL23pg4OP/z+Gc+f1IcA2O82Hl1tY3Nrfy24Wd3b39A/vwqKnDWFHWoKEIVdsnmgkesAZwEKwdKUakL1jLH9emfuueKc3D4A4mEetKMgz4gFMCRurZF56OJY5KHnDRZ8kkxY/YGxEwVJ5DLS1lgrESz5f4IS337KJTcWaFV8HNoIiyqvfsb68f0liyAKggWndcJ4JuQhRwKlha8GLNIkLHZMg6BgMime4ms/+l+MwofTwIlXkB4Jn6dyIhUuuJ9E2nJDDSy95U/M/rxDC46iY8iGJgAZ0vGsQCQ4inYeE+V4yCmBggVHFzK6YjoggFE+nCFl+mJhN3OYFVaJ5XXMO3TrF6naWTRyfoFJWQiy5RFd2gOmogip7QC3pFb9az9W59Wl/z1pyVzRyjhbJ+fgGNaqlW</latexit>

Noisy

[4] Patrini et al. CVPR 2017.

Incorporate

D

Conditional GAN [2, 3]

Generator:

Discriminator:

Baseline: cGAN

9

Dataset

D(x, y)<latexit sha1_base64="AYvJY9G/hMcRbQBxulga17SrgvU=">AAACA3icbZC7SgNBFIbPxluMt6ilzZAgRJSwa6NlUAvLCOYC2SXMTmaTITO768ysGJYtfQbbWNuJrQ9i6Zs4uRQm+sOBj/+cwzn8fsyZ0rb9ZeVWVtfWN/Kbha3tnd294v5BU0WJJLRBIh7Jto8V5SykDc00p+1YUix8Tlv+8HrSbz1SqVgU3utRTD2B+yELGMHaWN5NJXV9gZ6yMzQ66RbLdtWeCv0FZw7lWsk9HQNAvVv8dnsRSQQNNeFYqY5jx9pLsdSMcJoV3ETRGJMh7tOOwRALqrx0+nSGjo3TQ0EkTYUaTd3fGykWSo2EbyYF1gO13JuY//U6iQ4uvZSFcaJpSGaHgoQjHaFJAqjHJCWajwxgIpn5FZEBlphok9PCFV9kJhNnOYG/0DyvOobvTDhXMFMejqAEFXDgAmpwC3VoAIEHeIExvFrP1pv1bn3MRnPWfOcQFmR9/gA5Xpks</latexit><latexit sha1_base64="BA5RDs3afrtm4M93f27lvlOgNqE=">AAACA3icbZC7SgNBFIZn4y3GW9TSZkgQIkrYtdEyqIVlBHOB7BJmJ7PJkJnZdWZWXJYtfQPBVms7sfVBUvomTi6FSfzhwMd/zuEcfj9iVGnbHlm5ldW19Y38ZmFre2d3r7h/0FRhLDFp4JCFsu0jRRgVpKGpZqQdSYK4z0jLH16P+61HIhUNxb1OIuJx1Bc0oBhpY3k3ldT1OXzKzmBy0i2W7ao9EVwGZwblWsk9fRnVknq3+OP2QhxzIjRmSKmOY0faS5HUFDOSFdxYkQjhIeqTjkGBOFFeOnk6g8fG6cEglKaEhhP370aKuFIJ980kR3qgFntj879eJ9bBpZdSEcWaCDw9FMQM6hCOE4A9KgnWLDGAsKTmV4gHSCKsTU5zV3yemUycxQSWoXledQzfmXCuwFR5cARKoAIccAFq4BbUQQNg8ABewRt4t56tD+vT+pqO5qzZziGYk/X9C1mimrI=</latexit><latexit sha1_base64="LJoSMfVlXYI4/PbjbK8yMSWi4Xg=">AAACA3icbZBNS8NAEIYn9avWr6pHL4tFqCAl8aLHoh48VrAf0Iay2W7apbtJ3N2IIeTob/CqZ2/i1R/i0X/its3Btr4w8PDODDO8XsSZ0rb9bRVWVtfWN4qbpa3tnd298v5BS4WxJLRJQh7KjocV5SygTc00p51IUiw8Ttve+HrSbz9SqVgY3Oskoq7Aw4D5jGBtLPemmvY8gZ6yM5Sc9ssVu2ZPhZbByaECuRr98k9vEJJY0EATjpXqOnak3RRLzQinWakXKxphMsZD2jUYYEGVm06fztCJcQbID6WpQKOp+3cjxUKpRHhmUmA9Uou9iflfrxtr/9JNWRDFmgZkdsiPOdIhmiSABkxSonliABPJzK+IjLDERJuc5q54IjOZOIsJLEPrvOYYvrMr9as8nSIcwTFUwYELqMMtNKAJBB7gBV7hzXq23q0P63M2WrDynUOYk/X1Cw/Yl6M=</latexit>


[2] Mirza & Osindero. arXiv 2014. [3] Miyato & Koyama. ICLR 2018.

G

D

Conditional GAN [2, 3]

Generator:

Discriminator:

Limitation

G is optimized conditioned on noisy labels when trained with noisy labeled data.

Baseline: cGAN

10

Dataset

Conditionedon noisy label

Noisy

D(x, y)<latexit sha1_base64="AYvJY9G/hMcRbQBxulga17SrgvU=">AAACA3icbZC7SgNBFIbPxluMt6ilzZAgRJSwa6NlUAvLCOYC2SXMTmaTITO768ysGJYtfQbbWNuJrQ9i6Zs4uRQm+sOBj/+cwzn8fsyZ0rb9ZeVWVtfWN/Kbha3tnd294v5BU0WJJLRBIh7Jto8V5SykDc00p+1YUix8Tlv+8HrSbz1SqVgU3utRTD2B+yELGMHaWN5NJXV9gZ6yMzQ66RbLdtWeCv0FZw7lWsk9HQNAvVv8dnsRSQQNNeFYqY5jx9pLsdSMcJoV3ETRGJMh7tOOwRALqrx0+nSGjo3TQ0EkTYUaTd3fGykWSo2EbyYF1gO13JuY//U6iQ4uvZSFcaJpSGaHgoQjHaFJAqjHJCWajwxgIpn5FZEBlphok9PCFV9kJhNnOYG/0DyvOobvTDhXMFMejqAEFXDgAmpwC3VoAIEHeIExvFrP1pv1bn3MRnPWfOcQFmR9/gA5Xpks</latexit><latexit sha1_base64="BA5RDs3afrtm4M93f27lvlOgNqE=">AAACA3icbZC7SgNBFIZn4y3GW9TSZkgQIkrYtdEyqIVlBHOB7BJmJ7PJkJnZdWZWXJYtfQPBVms7sfVBUvomTi6FSfzhwMd/zuEcfj9iVGnbHlm5ldW19Y38ZmFre2d3r7h/0FRhLDFp4JCFsu0jRRgVpKGpZqQdSYK4z0jLH16P+61HIhUNxb1OIuJx1Bc0oBhpY3k3ldT1OXzKzmBy0i2W7ao9EVwGZwblWsk9fRnVknq3+OP2QhxzIjRmSKmOY0faS5HUFDOSFdxYkQjhIeqTjkGBOFFeOnk6g8fG6cEglKaEhhP370aKuFIJ980kR3qgFntj879eJ9bBpZdSEcWaCDw9FMQM6hCOE4A9KgnWLDGAsKTmV4gHSCKsTU5zV3yemUycxQSWoXledQzfmXCuwFR5cARKoAIccAFq4BbUQQNg8ABewRt4t56tD+vT+pqO5qzZziGYk/X9C1mimrI=</latexit><latexit sha1_base64="LJoSMfVlXYI4/PbjbK8yMSWi4Xg=">AAACA3icbZBNS8NAEIYn9avWr6pHL4tFqCAl8aLHoh48VrAf0Iay2W7apbtJ3N2IIeTob/CqZ2/i1R/i0X/its3Btr4w8PDODDO8XsSZ0rb9bRVWVtfWN4qbpa3tnd298v5BS4WxJLRJQh7KjocV5SygTc00p51IUiw8Ttve+HrSbz9SqVgY3Oskoq7Aw4D5jGBtLPemmvY8gZ6yM5Sc9ssVu2ZPhZbByaECuRr98k9vEJJY0EATjpXqOnak3RRLzQinWakXKxphMsZD2jUYYEGVm06fztCJcQbID6WpQKOp+3cjxUKpRHhmUmA9Uou9iflfrxtr/9JNWRDFmgZkdsiPOdIhmiSABkxSonliABPJzK+IjLDERJuc5q54IjOZOIsJLEPrvOYYvrMr9as8nSIcwTFUwYELqMMtNKAJBB7gBV7hzXq23q0P63M2WrDynUOYk/X1Cw/Yl6M=</latexit>


[2] Mirza & Osindero. arXiv 2014. [3] Miyato & Koyama. ICLR 2018.

G

Label-noise robust conditional GAN

Generator:

Discriminator: ( )

Solution

We correct the Dʼs input using the noise transition model.

yg ⇠ p(y|yg)<latexit sha1_base64="mYBu9DMzRUkQnQRSvMQ7XZ1Rx/E=">AAACIXicbZDLSsNAFIZPvNZ6q7p0MyiCIkjiRpdFNy4r2As0tUwm03ZwJgkzJ0KJeQZfQfAZ3OranbgTV76J0wuirQcGPv7/HM6ZP0ikMOi6H87M7Nz8wmJhqbi8srq2XtrYrJk41YxXWSxj3Qio4VJEvIoCJW8kmlMVSF4Pbs4Hfv2WayPi6Ar7CW8p2o1ERzCKVmqXDnwUMuRZP7/uEt8IRZL9H4ncEb9HcWgetEu77pE7LDIN3hh2yzv+4QMAVNqlLz+MWap4hExSY5qem2AroxoFkzwv+qnhCWU3tMubFiOquGllwy/lZM8qIenE2r4IyVD9PZFRZUxfBbZTUeyZSW8g/uc1U+yctjIRJSnyiI0WdVJJMCaDfEgoNGco+xYo08LeSliPasrQpvhnS6Bym4k3mcA01I6PPMuXNpwzGFUBtmEH9sGDEyjDBVSgCgzu4Qme4cV5dF6dN+d91DrjjGe24E85n983VKYU</latexit><latexit sha1_base64="N+GObZTuSN7epug/Sy8bqHdjL6w=">AAACIXicbZDLSgMxFIYzXmu9VV26CS1Ci1Bm3Oiy6MZlBXuBTi2ZNG1Dk5khOSMMY5/BjW/gM7jVtTtxV1z5JqbTIrb1QODj/8/hnPxeKLgG2x5bK6tr6xubma3s9s7u3n7u4LCug0hRVqOBCFTTI5oJ7rMacBCsGSpGpCdYwxteTfzGPVOaB/4txCFrS9L3eY9TAkbq5EoucNFlSTy662NXc4nD4q+EH7A7IJCapU6uYJfttPAyODMoVPLu6dO4Elc7uW+3G9BIMh+oIFq3HDuEdkIUcCrYKOtGmoWEDkmftQz6RDLdTtIvjfCJUbq4FyjzfMCp+nciIVLrWHqmUxIY6EVvIv7ntSLoXbQT7ocRMJ9OF/UigSHAk3xwlytGQcQGCFXc3IrpgChCwaQ4t8WTI5OJs5jAMtTPyo7hGxPOJZpWBh2jPCoiB52jCrpGVVRDFD2iF/SK3qxn6936sD6nrSvWbOYIzZX19QNXmKea</latexit><latexit sha1_base64="UmeAF4e1o7i8qOMDmJr5KPWFn0U=">AAACIXicbZDLSsNAFIYnXmu9VV26GSxCuymJG10W3bisYC/QxDKZTNuhM0mYORFC7DP4ED6DW127E3fiyjdxmhaxrQcGPv7/HM6Z348F12Dbn9bK6tr6xmZhq7i9s7u3Xzo4bOkoUZQ1aSQi1fGJZoKHrAkcBOvEihHpC9b2R1cTv33PlOZReAtpzDxJBiHvc0rASL1S1QUuApal47sBdjWXOK78SvgBu0MCuVntlcp2zc4LL4MzgzKaVaNX+naDiCaShUAF0brr2DF4GVHAqWDjoptoFhM6IgPWNRgSybSX5V8a41OjBLgfKfNCwLn6dyIjUutU+qZTEhjqRW8i/ud1E+hfeBkP4wRYSKeL+onAEOFJPjjgilEQqQFCFTe3YjokilAwKc5t8eXYZOIsJrAMrbOaY/jGLtcvZ+kU0DE6QRXkoHNUR9eogZqIokf0jF7Qq/VkvVnv1se0dcWazRyhubK+fgANzqSL</latexit>

D(x, y)<latexit sha1_base64="YMRI/FqFNl3NkGOT1mbxRivsoNY=">AAACDXicbZDLSsNAFIZP6q3WW1RcuRlahIoiiRtdFnXhsoK9QBPKZDpth84kYWYilpBn8BV0q2t34tZncOmbOL0sbOsPBz7+cw7n8AcxZ0o7zreVW1peWV3Lrxc2Nre2d+zdvbqKEklojUQ8ks0AK8pZSGuaaU6bsaRYBJw2gsH1qN94oFKxKLzXw5j6AvdC1mUEa2O17YObcuoFAj1mp8jTjHdoOsyO23bJOXPGQovgTqFUKXonzwBQbds/XiciiaChJhwr1XKdWPsplpoRTrOClygaYzLAPdoyGGJBlZ+O38/QkXE6qBtJU6FGY/fvRoqFUkMRmEmBdV/N90bmf71WoruXfsrCONE0JJND3YQjHaFRFqjDJCWaDw1gIpn5FZE+lphok9jMlUBkJhN3PoFFqJ+fuYbvTDhXMFEeDqEIZXDhAipwC1WoAYEUXuAV3qwn6936sD4nozlrurMPM7K+fgFj850T</latexit><latexit sha1_base64="jLL5zyj7xt0zLCtGYnYTckfnHm4=">AAACDXicbZDLSsNAFIYn9VbrLSqu3AwtQkUpiRtdFnXhsoK9QBPKZDJph84kYWYihpBncOMLuNW1O3HrM3Tpmzi9LGzrDwc+/nMO5/B7MaNSWdbIKKysrq1vFDdLW9s7u3vm/kFLRonApIkjFomOhyRhNCRNRRUjnVgQxD1G2t7wZtxvPxIhaRQ+qDQmLkf9kAYUI6Wtnnl0W80cj8On/Bw6ijKfZGl+2jMrVs2aCC6DPYNKveycvYzqaaNn/jh+hBNOQoUZkrJrW7FyMyQUxYzkJSeRJEZ4iPqkqzFEnEg3m7yfwxPt+DCIhK5QwYn7dyNDXMqUe3qSIzWQi72x+V+vm6jgys1oGCeKhHh6KEgYVBEcZwF9KghWLNWAsKD6V4gHSCCsdGJzVzye60zsxQSWoXVRszXf63CuwVRFcAzKoApscAnq4A40QBNgkIFX8AbejWfjw/g0vqajBWO2cwjmZHz/AoQ3npk=</latexit><latexit sha1_base64="OOqrchkajzgLhU2eG9CgclTlyDk=">AAACDXicbZDLSsNAFIYn9VbrLSqu3AwWoYKUxI0ui7pwWcFeoAllMpm0Q2cmYWYilpBn8Bnc6tqduPUZXPomTtssbOsPBz7+cw7n8AcJo0o7zrdVWlldW98ob1a2tnd29+z9g7aKU4lJC8cslt0AKcKoIC1NNSPdRBLEA0Y6wehm0u88EqloLB70OCE+RwNBI4qRNlbfPrqtZV7A4VN+Dj1NWUiycX7Wt6tO3ZkKLoNbQBUUavbtHy+MccqJ0JghpXquk2g/Q1JTzEhe8VJFEoRHaEB6BgXiRPnZ9P0cnhonhFEsTQkNp+7fjQxxpcY8MJMc6aFa7E3M/3q9VEdXfkZFkmoi8OxQlDKoYzjJAoZUEqzZ2ADCkppfIR4iibA2ic1dCXhuMnEXE1iG9kXdNXzvVBvXRTplcAxOQA244BI0wB1oghbAIAMv4BW8Wc/Wu/Vhfc5GS1axcwjmZH39Ajptm4o=</latexit>

G(z, yg)<latexit sha1_base64="SOIwORCvydCa/EjDTfWAMeO4m6Y=">AAACDXicbZC7SgNBFIbPxluMt1WxshkShIgSdm20DFpoGcFcIBvD7GSSDJnZXWZmhbjsM/gK2mptJ7Y+g6Vv4uRSmMQfDnz85xzO4fcjzpR2nG8rs7S8srqWXc9tbG5t79i7ezUVxpLQKgl5KBs+VpSzgFY105w2Ikmx8Dmt+4OrUb/+QKViYXCnhxFtCdwLWJcRrI3Vtg+ui4nnC/SYniKvj3UyTO97x2274JScsdAiuFMolPPeyTMAVNr2j9cJSSxooAnHSjVdJ9KtBEvNCKdpzosVjTAZ4B5tGgywoKqVjN9P0ZFxOqgbSlOBRmP370aChVJD4ZtJgXVfzfdG5n+9Zqy7F62EBVGsaUAmh7oxRzpEoyxQh0lKNB8awEQy8ysifSwx0SaxmSu+SE0m7nwCi1A7K7mGb004lzBRFg4hD0Vw4RzKcAMVqAKBBF7gFd6sJ+vd+rA+J6MZa7qzDzOyvn4BUw2dCA==</latexit><latexit sha1_base64="Xk5dcECSM836Sq211VVs1O96jYM=">AAACDXicbZC7SgNBFIZn4y3G26pY2QwJQkQJuzZaBi20jGAukF3D7GSSDJmZXWZmhXXZZ7DxBWy1thNbnyGlb+IksTCJPxz4+M85nMMfRIwq7TgjK7e0vLK6ll8vbGxube/Yu3sNFcYSkzoOWShbAVKEUUHqmmpGWpEkiAeMNIPh1bjffCBS0VDc6SQiPkd9QXsUI22sjn1wXU69gMPH7BR6A6TTJLvvH3fsklNxJoKL4P5CqVr0Tp5H1aTWsb+9bohjToTGDCnVdp1I+ymSmmJGsoIXKxIhPER90jYoECfKTyfvZ/DIOF3YC6UpoeHE/buRIq5UwgMzyZEeqPne2Pyv145178JPqYhiTQSeHurFDOoQjrOAXSoJ1iwxgLCk5leIB0girE1iM1cCnplM3PkEFqFxVnEN35pwLsFUeXAIiqAMXHAOquAG1EAdYJCCF/AK3qwn6936sD6noznrd2cfzMj6+gFzUZ6O</latexit><latexit sha1_base64="fziiGSVr2MA5GrP28d7XcL5PqqA=">AAACDXicbZDLSsNAFIYn9VbrLSqu3AwWoYKUxI0uiy50WcFeoIllMp20Q2cmYWYixJBn8Bnc6tqduPUZXPomTtssbPWHAx//OYdz+IOYUaUd58sqLS2vrK6V1ysbm1vbO/buXltFicSkhSMWyW6AFGFUkJammpFuLAniASOdYHw16XceiFQ0Enc6jYnP0VDQkGKkjdW3D65rmRdw+JifQm+EdJbm98OTvl116s5U8C+4BVRBoWbf/vYGEU44ERozpFTPdWLtZ0hqihnJK16iSIzwGA1Jz6BAnCg/m76fw2PjDGAYSVNCw6n7eyNDXKmUB2aSIz1Si72J+V+vl+jwws+oiBNNBJ4dChMGdQQnWcABlQRrlhpAWFLzK8QjJBHWJrG5KwHPTSbuYgJ/oX1Wdw3fOtXGZZFOGRyCI1ADLjgHDXADmqAFMMjAM3gBr9aT9Wa9Wx+z0ZJV7OyDOVmfPymHm38=</latexit>

Proposal: rcGAN

11

G

Dataset

D

Conditionedon clean label

Incorporate

Clean Noisy

Experiment: Comprehensive study

12

Experimental conditions

Dataset: CIFAR-10 [5], CIFAR-100 [5]

Noise: Symmetric [6], Asymmetric [4] (Noise rate ∈ {0, 0.1, 0.3, 0.5, 0.7, 0.9})

GAN configurations: DCGAN [7], WGAN-GP [8], CT-GAN [9], SN-GAN [10]

Comparison: AC-GAN vs. rAC-GAN, cGAN vs. rcGANEvaluation metrics: FID [11], Intra FID [3], GAN-test [12], GAN-train [12]

We tested 336 conditions in total.

[5] Krizhevsky. 2009. [6] van Rooyen et al. NIPS 2015. [4] Patrini et al. CVPR 2017.[7] Radford et al. ICLR 2016. [8] Gulrajani et al. NIPS 2017. [9] Wei et al. ICLR 2018. [10] Miyato et al. ICLR 2018.

[11] Heusel et al. NIPS 2017. [3] Miyato & Koyama. ICLR 2018. [12] Shmelkov et al. ECCV 2018.

Experiment: Quantitative results

13

Comparison between the proposed model and the baselines across all the conditions

Baseline

Prop

osed

rAC-GAN vs. AC-GANrcGAN vs. cGAN

Larger is better

Baseline

Prop

osed

rAC-GAN vs. AC-GANrcGAN vs. cGAN

Larger is better

(a) GAN-test (b) GAN-train

The proposed models outperform the baselines in most cases.

Experiment: Qualitative results I

14

CIFAR-10 symmetric noise (uniform noise)

rAC-CT-GAN (30.2)AC-CT-GAN (36.4)

rcSN-GAN (28.6)cSN-GAN (72.0)

Baseline Proposed

Significantlydegraded

The number indicates Intra FID. A smaller value is better.

Experiment: Qualitative results II

15

CIFAR-10 asymmetric noise (class-dependent noise, e.g., cat ⇄ dog)

rAC-CT-GAN (27.2)AC-CT-GAN (45.7)

rcSN-GAN (25.6)cSN-GAN (33.7)

Baseline Proposed

The number indicates Intra FID. A smaller value is better.

Confusebetweenflipped classes

Cat Dog Cat Dog

Cat Dog Cat Dog

Further analyses

16

Effects of estimated noise transition modelWe examined the effect whenthe noise transition model is estimated from data [4].

Evaluation of improved techniqueWe validated the effect of an improved technique,which we developed to boost the performancein severely noisy setting.

Evaluation on real-world noiseWe tested on Clothing1M [13],which includes real-world noisy labeled data.

LMI = Ez⇠p(z),yg⇠p(y)

[� logQ(y = yg|G(z, yg))]<latexit sha1_base64="r7Lei40m4T1XLN+K2euvI83CLpg=">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</latexit><latexit sha1_base64="r7Lei40m4T1XLN+K2euvI83CLpg=">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</latexit><latexit sha1_base64="r7Lei40m4T1XLN+K2euvI83CLpg=">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</latexit>

¯

x

i= argmax

x2X 0C 0

(y = i|x)

T 0i,j = C 0

(y = j|¯xi)

<latexit sha1_base64="mlt41Tm3wmyCbhhigODxdDwr2B4=">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</latexit><latexit sha1_base64="mlt41Tm3wmyCbhhigODxdDwr2B4=">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</latexit><latexit sha1_base64="mlt41Tm3wmyCbhhigODxdDwr2B4=">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</latexit>

Robust two-step training algorithm [4]

Loss for improving the performance

[4] Patrini et al. CVPR 2017. [13] Xiao et al. CVPR 2015.

Qualitative results on Clothing1M

Thank you!

17

Our code is publicly available at https://github.com/takuhirok/rGAN/

PosterSession: Synthesis

#: 133Time: Tuesday 15:20-18:00

DDD/CD/C

AC-GAN [1] rAC-GAN cGAN [2, 3] rcGAN


label-noise robust generative adversarial networks · label-noise robust generative adversarial...

Documents