Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI

Generative Adversarial Networks (GANs) - Ian Goodfellow, OpenAI

• 1.
  Generative Adversarial Networks (GANs) IanGoodfellow, OpenAI Research Scientist Presentation at AI With the Best, 2016-09-24
• 2.
  (Goodfellow 2016) Generative Modeling •Density estimation • Sample generation Training examples Model samples
• 3.
  (Goodfellow 2016) Adversarial NetsFramework Input noise Z Differentiable function G x sampled from model Differentiable function D D tries to output 0 x sampled from data Differentiable function D D tries to output 1
• 4.
  (Goodfellow 2016) DCGAN Architecture (Radfordet al 2015) Most “deconvs” are batch normalized
• 5.
  (Goodfellow 2016) DCGANs forLSUN Bedrooms (Radford et al 2015)
• 6.
  (Goodfellow 2016) Vector SpaceArithmeticCHAPTER 15. REPRESENTATION LEARNING - + = Figure 15.9: A generative model has learned a distributed representation that disentangl he concept of gender from the concept of wearing glasses. If we begin with the repr entation of the concept of a man with glasses, then subtract the vector representing th oncept of a man without glasses, and finally add the vector representing the conce f a woman without glasses, we obtain the vector representing the concept of a woma Man with glasses Man Woman Woman with Glasses
• 7.
  (Goodfellow 2016) Mode Collapse •Fully optimizing the discriminator with the generator held constant is safe • Fully optimizing the generator with the discriminator held constant results in mapping all points to the argmax of the discriminator • Can partially fix this by adding nearest-neighbor features constructed from the current minibatch to the discriminator (“minibatch GAN”) (Salimans et al 2016)
• 8.
  (Goodfellow 2016) Minibatch GANon CIFAR Training Data Samples (Salimans et al 2016)
• 9.
  (Goodfellow 2016) Minibatch GANon ImageNet (Salimans et al 2016)
• 10.
  (Goodfellow 2016) Cherry-Picked Results
• 11.
  (Goodfellow 2016) Text toImage with GANs n Yan, Lajanugen Logeswaran REEDSCOT1 , AKATA2 , XCYAN1 , LLAJAN1 SCHIELE2 ,HONGLAK1 r, MI, USA (UMICH.EDU) cs, Saarbr¨ucken, Germany (MPI-INF.MPG.DE) images from text l, but current AI oal. However, in ful recurrent neu- been developed ature representa- utional generative ave begun to gen- s of specific cat- covers, and room elop a novel deep ion to effectively nd image model- s from characters capability of our ages of birds and iptions. this small bird has a pink breast and crown, and black primaries and secondaries. the flower has petals that are bright pinkish purple with white stigma this magnificent fellow is almost all black with a red crest, and white cheek patch. this white and yellow flower have thin white petals and a round yellow stamen Figure 1. Examples of generated images from text descriptions.(Reed et al 2016)
• 12.
  (Goodfellow 2016) Generating Pokémon youtube (YotaIshida)
• 13.
  (Goodfellow 2016) Single ImageSuper-Resolution (Ledig et al 2016)
• 14.
  (Goodfellow 2016) iGAN youtube (Zhu etal 2016)
• 15.
  (Goodfellow 2016) Introspective Adversarial Networks youtube
• 16.
  (Goodfellow 2016) Conclusion • GANsare generative models based on supervised learning and game theory • GANs learn to generate realistic samples • Like other generative models, GANs still need a lot of improvement