A u t o E n c o d e r s & G e n e r a t i v e
a d v e r s a r i a l m o d e l s
U n s u p e r v i s e d l e a r n i n g
G e n e r a t i v e m o d e l s c a p a b i l i t i e s
Dr. Haithem Hermessi
•
•
Sr. Comp u ter Vision En g in eer @ SCYLLA
AI research Scientist @ LIMTIC - University of Tunis El Manar
AutoEncoders: why?
Image generation
Interpolation in Pixel
Space
Super resolution
Transformation
Autoencoders
Unsupervised learning
Autoencoder
Decoder
Encoder
If “tight weights”,
then
Reconstruction losses
binary
input
real valued
input
Publicité
Under-/over-complete hidden layer
Denoising autoencoder
corruption
denoising
We assume we are injecting the same noisy distribution
we’re going to observe in reality. In this way, we can learn how
to robustly recover from it.
Contractive autoencoder
penalises insensitivity to
reconstruction directions
penalises
sensitivity to the
any direction
penalised
direction
penalised &
incentivised
direction
Basic auto-encoder
How can we “go back”:
learn the distribution
-
or
- enforce some
structure
Dealing with distributions
Generative models
Auto-encoder (recap)
Publicité
Decoder
Encoder
Variational auto-encoder
Classic auto-encoder
Decoder
Sampler
Encoder
Decoder
Encoder
Variational auto-
encoder
Decoder
Sampler
Encoder
reconstruction loss
Variational auto-
encoder
encoder + noise
decode
r
To go from the latent to
the input space we need
to:
•
learn the distribution
or
• enforce some
structure
Publicité
Variational auto-encoder
Variational auto-encoder
Generative adversarial nets
Unsupervised learning / Generative models
Variational auto-
encoder
Generative adversarial
network
Decode
r
Sampl
er
Encod
er
Cos
t
Generato
r
Sampl
er
Generative adversarial
network
C
o
s
t
Generato
r
Publicité
Sampl
er
Training
Possible choice of
C(x):
Zhao, Mathieu, LeCun (2016) Energy-based generative adversarial
network
Generative adversarial
network
high
cost
generato
r
low
cost
Major pitfalls
• Vanishing gradients
• Mode collapse
• Unstable
convergence
Reference:
• Nivdia Deep learning teaching Kit
• Stanford University course CS231n: Convolutional Neural
Networks for Visual Recognition
• University of Maryland course CMSC498L
26