Lecture 8.

Generative Adversarial Network

 GAN was first introduced by Ian Goodfellow et al in 2014
 Have been used in generating images, videos, poems, some simple
conversation.
 Note, image processing is easy (all animals can do it), NLP is hard (only
human can do it).
 This co-evolution approach might have far-reaching implications. Bengio:
this may hold the key to making computers a lot more intelligent.

 Ian Goodfellow:
https://www.youtube.com/watch?v=YpdP_0-IEOw
 Radford, (generate voices also here)
https://www.youtube.com/watch?v=KeJINHjyzOU
 Tips for training GAN: https://github.com/soumith/ganhacks
Autoencoder

As close as possible

code
NN NN
Encoder Decoder
code

Randomly NN
generate a vector Decoder Image ?
as code
Autoencoder with 3 fully connected layers
Training: model.fit(X,X)
Cost function: Σk=1..N (xk – x’k)2

Large  small, learn to compress

 Auto-encoder
2D code
NN
code
Decoder
NN
Decoder

-1.5 1.5

NN
Decoder
Auto-encoder

-1.5 1.5
 Auto-encoder

NN NN
input Encoder
output
Decoder

code
VAE Minimize
m1
m2 reconstruction error
NN m3 c1
input NN
Encoder σ1 exp + c2 output
Decoder
σ2 c3
σ3
X ci = exp(σi)ei + mi
From a normal e1
distribution e2 Minimize
e3

Auto-Encoding Variational Bayes, Σi=1..3 [exp(σi)−(1+σi)+(mi)2 ]

https://arxiv.org/abs/1312.6114 This constrains σi approacing 0 is good
 Problems of VAE
 It does not really try to simulate real images

code
NN
Output As close as
Decoder
possible

One pixel difference to Also one pixel

the target difference to the target

Realistic Fake
VAE treats these the same
 Gradual and step-wise generation

NN NN NN
Generator Generator Generator
v1 v2 v3

Generated Discri- Discri- Discri-

images minator minator minator
v1 v2 v3

These are Real images:

Binary classifiers
GAN – Learn a discriminator

Randomly NN
Generator
sample a v1
vector 0 0 0 0

Something like Real images

Decoder in VAE Sampled from
DB: 1 1 1 1

Discri-
image minator 1/0 (real or fake)
v1
 Randomly sample
GAN – Learn a generator a vector

Train NN
Updating the parameters of this Generator
generator v2 v1

The output be classified They have

as “real” (as close to 1 Opposite
objectives
as possible)

Generator + Discriminator =
a network Do not Discri-
Train
minator
Using gradient descent to This
v1
update the parameters in the
generator, but fix the
discriminator 1.0 0.13
 Generating 2nd element figures

You can use the following to start a project (but this is in Chinese):
Source of images: https://zhuanlan.zhihu.com/p/24767059
From Dr. HY Lee’s notes.
DCGAN: https://github.com/carpedm20/DCGAN-tensorflow
 GAN – generating 2nd element figures

100 rounds

This is fast, I think you can use your CPU

GAN – generating 2nd element figures

1000 rounds
GAN – generating 2nd element figures

2000 rounds
GAN – generating 2nd element figures

5000 rounds
GAN – generating 2nd element figures

10,000 rounds
GAN – generating 2nd element figures

20,000 rounds
GAN – generating 2nd element figures

50,000 rounds
Next few images from Goodfellow lecture

Traditional mean-squared
Error, averaged, blurry
Last 2 are by deep learning approaches.
Similar to word embedding (DCGAN paper)
256x256 high resolution pictures
by Plug and Play generative network
From natural language to pictures
Deriving GAN

 During the rest of this lecture, we will go

thru the original ideas and derive GAN.
 I will avoid the continuous case and stick
to simple explanations.
 Maximum Likelihood Estimation
 Give a data distribution Pdata(x)
 We use a distribution PG(x;θ) parameterized by θ to
approximate it
 E.g. PG(x;θ) is a Gaussian Mixture Model, where θ contains
means and variances of the Gaussians.
 We wish to find θ s.t. PG(x;θ) is close to Pdata(x)
 In order to do this, we can sample
{x1,x2, … xm} from Pdata(x)
 The likelihood of generating these
xi’s under PG is
L= Πi=1…m PG(xi; θ)
 Then we can find θ* maximizing the L.
KL (Kullback-Leibler) divergence
 Discrete:
DKL(P||Q) = ΣiP(i)log[P(i)/Q(i)]
 Continuous:
∞
DKL(P||Q) = p(x)log [p(x)/q(x)]
−∞
 Explanations:
Entropy: - ΣiP(i)logP(i) - expected code length (also optimal)
Cross Entropy: - ΣiP(i)log Q(i) – expected coding
length using optimal code for Q
DKL= ΣiP(i)log[P(i)/Q(i)] = ΣiP(i)[logP(i) – logQ(i)], extra bits
JSD(P||Q) = ½ DKL(P||M)+ ½ DKL(Q||M), M= ½ (P+Q), symmetric KL
* JSD = Jensen-Shannon Divergency
Maximum Likelihood Estimation
θ* = arg maxθ Πi=1..mPG(xi; θ) 
arg maxθ log Πi=1..mPG(xi; θ)
= arg maxθ Σi=1..m log PG(xi; θ), {x1,..., xm} sampled from Pdata(x)
= arg maxθ Σi=1..m Pdata(xi) log PG(xi; θ) --- this is cross entropy
≅ arg maxθ Σi=1..m Pdata(xi) log PG(xi; θ) - Σi=1..m Pdata(xi )logPdata(x i)
= arg minθ KL (Pdata(x) || PG(x; θ)) --- this is KL divergence

Note: PG is Gaussian mixture model, finding best θ will still be Gaussians, this
only can generate a few blubs. Thus this above maximum likelihood approach
does not work well.

Next we will introduce GAN that will change PG, not just estimating PG is
parameters We will find best PG , which is more complicated and structured, to
approximate Pdata.
 Thus let’s use an NN as PG(x; θ)
PG(x,θ) Pdata(x)

θ
Prior
distribution Smaller
of z dimension
Larger How to compute the
dimension
likelihood?
PG(x) = Integrationz Pprior(z) I[G(z)=x]dz https://blog.openai.com/generative-models/
Basic Idea of GAN

 Generator G Hard to learn PG by maximum likelihood

 G is a function, input z, output x
 Given a prior distribution Pprior(z), a probability distribution PG(x) is
defined by function G
 Discriminator D
 D is a function, input x, output scalar
 Evaluate the “difference” between PG(x) and Pdata(x)
 In order for D to find difference between P data from PG, we
need a cost function V(G,D):
G*=arg minGmaxDV(G,D)
Note, we are changing distribution G, not just update
its parameters (as in the max likelihood case).
Basic Idea G* = arg minGmaxD V(G,D)

Pick JSD function: V = Ex~P_data [log D(x)] + Ex~P_G[log(1-D(x))]

Given a generator G, maxDV(G,D) evaluates the

“difference” between PG and Pdata
Pick the G s.t. PG is most similar to Pdata

V(G1,D) V(G2,D) V(G3,D)

G1 G2 G3
MaxDV(G,D), G*=arg minGmaxDV(G,D)

 Given G, what is the optimal D* maximizing

V = Ex~P_data [log D(x)] + Ex~P_G[log(1-D(x))]
= Σ [ Pdata(x) log D(x) + PG(x) log(1-D(x) ]

Thus: D*(x) = Pdata(x) / (Pdata(x)+PG(x))

Assuming D(x) can have any value here

 Given x, the optimal D* maximizing is:
f(D) = alogD + blog(1-D)  D*=a/(a+b)
 maxDV(G,D), G* = arg minGmaxD V(G,D)

D1*(x) = Pdata(x) / (Pdata(x)+PG_1(x))

D2*(x) = Pdata(x) / (Pdata(x)+PG_2(x))

“difference” between
PG1 and Pdata

V(G1,D*1)

V(G1,D) V(G2,D) V(G3,D)

maxDV(G,D) V = Ex~P_data [log D(x)]
+ Ex~P_G[log(1-D(x))]

maxD V(G,D)
= V(G,D*), where D*(x) = Pdata / (Pdata + PG), and
1-D*(x) = PG / (Pdata + PG)
= Ex~P_data log D*(x) + Ex~P_G log (1-D*(x))
≈ Σ [ Pdata (x) log D*(x) + PG(x) log (1-D*(x)) ]
= -2log2 + 2 JSD(Pdata || PG ),

JSD(P||Q) = Jensen-Shannon divergence

= ½ DKL(P||M)+ ½ DKL(Q||M)
where M= ½ (P+Q).
DKL(P||Q) = Σ P(x) log P(x) /Q(x)
Summary: V = Ex~P_data [log D(x)]
+ Ex~P_G[log(1-D(x))]
 Generator G, Discriminator D
 Looking for G* such that

G* = arg minGmaxD V(G,D)

 Given G, maxD V(G,D)

= -2log2 + 2JSD(Pdata(x) || PG(x))
 What is the optimal G? It is G that makes JSD
smallest = 0:
PG(x) = Pdata (x)
Algorithm G* = arg minGmaxD V(G,D)
L(G), this is the
loss function
 To find the best G minimizing the loss function L(G):
θG  θG =−η L(G)/ θG , θG defines G
 Solved by gradient descent. Having max ok. Consider
simple case: If Di(x) is the
f(x) = max {D1(x), D2 （ x), D3(x)} Max in that region,
then do dDi(x)/dx

D1(x) D3(x)
D2(x)

dD1(x)/dx dD2(x)/dx dD3(x)/dx

 G* = arg minGmaxD V(G,D)
Algorithm L(G)

 Given G0
 Find D*0 maximizing V(G0,D)
V(G0,D0*) is the JS divergence between Pdata(x) and PG0(x)
 θG  θG −η ΔV(G,D0*) / θG  Obtaining G1 (decrease JSD)
 Find D1* maximizing V(G1,D)
V(G1,D1*) is the JS divergence between Pdata(x) and PG1(x)
 θG  θG −η ΔV(G,D1*) / θG  Obtaining G2 (decrease JSD)
 And so on …
In practice … V = Ex~P_data [log D(x)]
+ Ex~P_G[log(1-D(x))]

 Given G, how to compute maxDV(G,D)?

 Sample {x1, … ,xm} from Pdata
 Sample {x*1, … ,x*m} from generator PG
Maximize:
V’ = 1/m Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))
Positive example Negative example
D must accept D must reject
This is what a Binary Classifier do
Output is D(x) Minimize Cross-entropy
If x is a positive example Minimize –log D(x)
If x is a negative example Minimize –log(1-D(x))
 Binary Classifier
Output is f(x) Minimize Cross-entropy
If x is a positive example Minimize –log f(x)
If x is a negative example Minimize –log(1-f(x))

D is a binary classifier (can be deep) with parameters θd

{x1,x2, … xm} from Pdata (x) Positive examples

{x*1,x*2, … x*m} from PG(x) Negative examples

Minimize L = - V’

or

Maximize V’ = Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))

 Initialize θd for D and θg for G
Algorithm
Can only find lower
bound of JSD or
 In each training iteration maxDV(G,D)
Sample m examples {x1,x2, … xm} from data distribution Pdata(x)
Ian Goodfellow
comment: this  Sample m noise samples {z , … , z } from a simple prior Pprior(z)
1 m

is also done once

 Obtain generated data {x*1, … , x*m}, x*i=G(zi)
Learning D Update discriminator parameters θd to maximize
 V’ ≈ 1/m Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))
 θd  θd + ηΔV’(θd) (gradient ascent)
Repeat  Simple another m noise samples {z1,z2, … zm} from the prior
k times Pprior(z) ， G(zi)=x*i
 Update generator parameters θg to minimize
V’= 1/mΣi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))
θg  θg − ηΔV’(θg) (gradient descent)
Learning G

Only
Once
Objective Function for Generator
in Real Implementation

V = Ex~P_data [log D(x)

+ Ex~P_G[log(1-D(x))]

Training slow at the beginning

V = Ex~P_G [ − log (D(x)) ]

Real implementation:
label x from PG as positive
Some issues in training GAN

M. Arjovsky, L. Bottou, Towards principled

methods for training generative adversarial
networks, 2017.
Evaluating JS divergence
Discriminator is too strong: for all three
Generators, JSD = 0

Martin Arjovsky, Léon Bottou, Towards Principled Methods for Training

Generative Adversarial Networks, 2017, arXiv preprint
Evaluating JS divergence https://arxiv.org/a
bs/1701.07875

 JS divergence estimated by discriminator

telling little information

Weak Generator Strong Generator

 Discriminator
1 for all positive examples 0 for all negative examples

V = Ex~P_data [log D(x)] + Ex~P_G[log(1-D(x))]

= 1/m Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))

maxDV(G,D) = -2log2 + 2 JSD(Pdata || PG ) =0

log 2 when Pdata and PG differ
completely

Reason 1. Approximate by sampling

Weaken your discriminator?

Can weak discriminator

compute JS divergence?
 Discriminator
GAN implementation
1 0 estimation
V = Ex~P_data [log D(x)] + Ex~P_G[log(1-D(x))]
= 1/m Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i)) ≈0

maxDV(G,D) = -2log2 + 2 JSD(Pdata || PG ) = 0

Theoretical estimation log2

Reason 2. the nature of data

Pdata(x) and PG(x) have very little

overlap in high dimensional space
Evolution http://www.guokr.com/post/773890/

Better
Evolution needs to be smooth:

JSD(PG_0 || Pdata) = log2

PG_0(x) Pdata(x)

……
Better PG_50(x) Pdata(x) JSD(PG_50 || Pdata) = log2

Not really better

……

……
PG_100(x)
Pdata(x) JSD(PG_100 || Pdata) = 0
One simple solution: add noise

 Add some artificial noise to the inputs of

discriminator
 Make the labels noisy for the discriminator
Discriminator cannot perfectly separate real and generated
data

Pdata(x) and PG(x) have

some overlap

Noises need to decay over time

 Mode Collapse
Converge to same faces
Generated
Distribution

Data
Distribution

Sometimes, this is hard to tell since

one sees only what’s generated, but not what’s missed.
 Mode Collapse Example
Pdata
8 Gaussian distributions:

What we want
…

In reality …
Text to Image, by conditional GAN
 "red flower with
Text to Image black center"
- Results From CY Lee lecture

Project topic: Code and data are all on web, many possibilities!
 Algorithm WGAN
 In each training iteration
Sample m examples {x1,x2, … xm} from data distribution Pdata(x)
Ian Goodfellow
comment: this  Sample m noise samples {z , … , z } from a simple prior Pprior(z)
1 m

is also done once

 Obtain generated data {x*1, … , x*m}, x*i=G(zi)
Learning D Update discriminator parameters θd to maximize
 V’ ≈ Σi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))
 θd  θd + ηΔV’(θd) (gradient ascent plus weight clipping)
Repeat Simple another m noise samples {z1,z2, … zm} from the prior
k times Pprior(z) ， G(zi)=x*i
 Update generator parameters θg to minimize
V’= 1/mΣi=1..m logD(xi) + 1/m Σi=1..m log(1-D(x*i))
θg  θg − ηΔV’(θg) (gradient descent)
Learning G

Only
Once
Experimental Results

 Approximate a mixture of Gaussians by

single mixture
WGAN Background

 We have seen that JSD does not give

GAN a smooth and continuous
improvement curve.
 We would like to find another distance
which gives that.
 This is the Wasserstein Distance or earth
mover’s distance.
Earth Mover’s Distance
 Considering one distribution P as a pile of earth (total
amount of earth is 1), and another distribution Q (another
pile of earth) as the target
 The “earth mover’s distance” or “Wasserstein Distance”
is the average distance the earth mover has to move the
earth in an optimal plan.

d
Earth Mover’s Distance: best plan to
move
P

Q
 JS vs Earth Mover’s Distance

d0 d50 d100

PG_0 Pdata …… PG_50 Pdata …… PG_100 Pdata

JS(PG_0, Pdata) = log2 JS(PG_50, Pdata) = log2 JS(PG_100, Pdata) = 0

W(PG_0, Pdata)=d0 W(PG_50, Pdata)=d50 W(PG_100, Pdata)=0

 Explaining WGAN
 Let W be the Wasserstein distance.
W(Pdata, PG) = maxD is 1-Lipschitz[Ex~P_data D(x) – Ex~P_G D(x)]

Where a function f is a Blue: D(x) for original GAN

k-Lipschitz function if Green: D(x) for WGAN

||f(x1) – f(x2) ≤ k||x1 – x2 ||

How to guarantee this?

Weight clipping: for all
parameter updates, if w>c
Then w=c, if w<-c, then w=-c. WGAN will provide gradient to
push PG towards Pdata
Earth Mover Distance Examples:
Multi-layer perceptron