Generative Adversarial Networks Ian Goodfellow Research Scientist - - PowerPoint PPT Presentation
Generative Adversarial Networks Ian Goodfellow Research Scientist GPU Technology Conference San Jose, California 2016-04-05 Generative Modeling Have training examples: x p train ( x ) Want a model that can draw samples: x p
Ian Goodfellow Research Scientist
GPU Technology Conference San Jose, California 2016-04-05
x ∼ ptrain(x) x ∼ pmodel(x) pmodel(x) = pdata(x) (Images from Toronto Face Database)
x Probability Density
q∗ = argminqDKL(pkq) p(x) q∗(x)
x Probability Density
q∗ = argminqDKL(qkp) p(x) q∗(x)
Put high probability where there should be high probability Put low probability where there should be low probability (Deep Learning, Goodfellow, Bengio, and Courville 2016)
Input noise Z Differentiable function G x sampled from model Differentiable function D D tries to
x sampled from data Differentiable function D D tries to
x x z
(“Generative Adversarial Networks”, Goodfellow et al 2014)
Data distribution Model distribution
Optimal D(x) for any pdata(x) and pmodel(x) is always
...
Poorly fit model After updating D After updating G Mixed strategy equilibrium Data distribution Model distribution
Discriminator response
MNIST digit dataset Toronto Face Dataset (TFD)
(Alec Radford)
(Denton+Chintala et al 2015)
(Denton+Chintala et al 2015)
(Radford et al 2015)
=
(Radford et al 2015) Man wearing glasses Man Woman Woman wearing glasses
Input Reconstruction
(Chelsea Finn)
(Lotter et al, 2015) Capture predictable details regardless of scale
assigning high probability to all samples
function
even if they are small or faint