9953056974 Young Call Girls In Mahavir enclave Indian Quality Escort service
PixelGAN Autoencoders
1. Alireza Makhzani PixelGAN Autoencoders / 27
CIFAR Deep Learning Summer School
Montreal, Canada
Machine learning Group
University of Toronto
PixelGAN Autoencoders
Alireza Makhzani, Brendan Frey
June 29th, 2017
1
3. Alireza Makhzani PixelGAN Autoencoders / 27
PixelCNNs
33
✦ Learn the image statistics directly at the pixel level.
✦ Good at modelling low-level pixel statistics.
✦ Samples lack global structure.
✦ Lacking latent representation.
✦ Conditional PixelCNNs can learn conditional densities.
van den Oord et al., 2016
4. Alireza Makhzani PixelGAN Autoencoders / 27
Variational Autoencoders
4
log p(x) > Eq(z|x)[ log(p(x|z)] KL(q(z|x)kp(z))
4
Kingma et al., 2013
✦ Good at capturing the global structure, but samples are blurry.
✦ Learn hierarchical representations useful for down-stream tasks.
✦ Attempts at combining PixelCNN with VAEs:
• PixelVAE (Gulrajani et al., 2016)
• VLAE (Chen et al., 2017)
5. Alireza Makhzani PixelGAN Autoencoders / 27
Adversarial Autoencoders
5
A
B
C
D
Code Space
of MNIST:
Gaussian Prior Mixture of Gaussians
Makhzani et al., 2015
7. Alireza Makhzani PixelGAN Autoencoders / 27
Limitations of Variational/Adversarial Autoencoders
7
✦ All the image statistics are captured by the single latent vector.
label, style
global and local
VAE
None
z
x
p(z)
p(x|z)
Latent Variable
Deterministic
(factorized Gaussians)
8. Alireza Makhzani PixelGAN Autoencoders / 27
PixelGAN Autoencoders
8
✦ The image statistics are captured jointly by the latent vector and the
autoregressive decoder.
PixelCNN
Latent Variablez
x
p(z)
p(x|z)
9. Alireza Makhzani PixelGAN Autoencoders / 27
PixelGAN Autoencoders
9
PixelCNN
Latent Variablez
x
p(z)
p(x|z)
GAN
✦ The image statistics are captured jointly by the latent vector and the
autoregressive decoder.
10. Alireza Makhzani PixelGAN Autoencoders / 27
PixelGAN Autoencoders
10
PixelGAN
(Gaussian)
Global
(low-frequency)
Local
(high-frequency)
PixelGAN
(Categorical)
Discrete
(label)
Continuous
(Style)
PixelCNN
Latent Variablez
x
p(z)
p(x|z)
GAN
✦ The image statistics are captured jointly by the latent vector and the
autoregressive decoder.
11. Alireza Makhzani PixelGAN Autoencoders / 27
PixelGAN Autoencoders
11
Semi-supervised Learning
PixelCNN
Latent Variablez
x
p(z)
p(x|z)
GAN
PixelGAN
(Gaussian)
Global
(low-frequency)
Local
(high-frequency)
PixelGAN
(Categorical)
Discrete
(label)
Continuous
(Style)
✦ The image statistics are captured jointly by the latent vector and the
autoregressive decoder.
12. Alireza Makhzani PixelGAN Autoencoders / 27
PixelGAN Autoencoders
12
Cost function of PixelGAN = Reconstruction + Adversarial Cost