Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018

http://bit.ly/dlai2018
Santiago Pascual de la Puente
santi.pascual@upc.edu
PhD Candidate
Universitat Politecnica de Catalunya
Technical University of Catalonia
Deep Generative Models I
Variational Autoencoders
#DLUPC

Outline
● What is in here?
● Introduction
● Taxonomy
● Variational Auto-Encoders (VAEs) (DGMs I)
● Generative Adversarial Networks (GANs)
● PixelCNN/Wavenet
● Normalizing Flows (and flow-based gen. models)
○ Real NVP
○ GLOW
● Models comparison
● Conclusions
2

Disclaimer
4
● We are going to make a shift in the modeling paradigm:
discriminative → generative.
● Is this a type of neural network? No → either fully connected
networks, convolutional networks or recurrent networks fit,
depending on how we condition data points (sequentially,
globally, etc.).
● This is a very fun topic with which we can make a network paint,
sing or write.

Planning for our generative trip
5
● 5/11/2018 → DGM I: Introduction to gen models + VAEs
○ (... in b/w ...) Methodology + RNN lessons
● 19/11/2018 → DGM II: GANs
● 19/11/2018 → DGM III: likelihood models (pixelCNN + flow models)
● 26/11/2018 → Practical lesson on DGMs (code day)

What we are used to do with Neural Nets
7
0
1
0
input
Network
output
class
Figure credit: Javier Ruiz
Discriminative model → aka. tell me the probability of some ‘Y’ responses given ‘X’ inputs. Here we
don’t care about the process that generated ‘X’. we just detect some patterns in the input to give
an answer. This gives us some outcomes probabilities given some ‘X’ data: P(Y | X).
P(Y = [0,1,0] | X = [pixel1
, pixel2
, …, pixel784
])

What is a generative model?
8
We have datapoints that emerge from some
generating process (landscapes in the nature,
speech from people, etc.)
X = {x1
, x2
, …, xN
}
Having our dataset X with example datapoints
(images, waveforms, written text, etc.) each point xn
lays in some M-dimensional space.

9
We have datapoints that emerge from some
generating process (landscapes in the nature,
speech from people, etc.)
X = {x1
, x2
, …, xN
}
Having our dataset X with example datapoints
(images, waveforms, written text, etc.) each point xn
lays in some M-dimensional space.
Each point xn
comes from an M-dimensional
probability distribution P(X) → Model it!

10
We have datapoints that emerge from some generating
process (landscapes in the nature, speech from people,
etc.)
X = {x1
, x2
, …, xN
}
Y = {y1
, y2
, …, yN
}
Having our dataset X with example datapoints (images,
waveforms, written text, etc.) each point xn
lays in some
M-dimensional space. Each point yn
lays in some
K-dimensional space. M can be different than K.
DOG
CAT
TRUCK
PIZZA
THRILLER
SCI-FI
HISTORY
/aa/
/e/
/o/
We can also modeled joint probabilities to apply
conditioning variables on the generative
process: P(X, Y)

11
1) We want our model with parameters θ={weights, biases} to output samples
distributed Pmodel, matching the distribution of our training data Pdata or P(X).
2) We can sample points from Pmodel plausibly looking Pdata distributed.

12
1) We want our model with parameters θ={weights, biases} to output samples
distributed Pmodel, matching the distribution of our training data Pdata or
P(X).
2) We can sample points from Pmodel plausibly looking Pdata distributed.
We have not mentioned any network structure
or model input data format, just a requirement
on what we want in the output of our model.

13
We can generate any type of data, like speech waveform samples or image pixels.
M samples = M dimensions
32
32
.
.
.
Scan and unroll
x3 channels
M = 32x32x3 = 3072 dimensions

Our learned model should be able to make up new samples from the
distribution, not just copy and paste existing samples!
14
Figure from NIPS 2016 Tutorial: Generative Adversarial Networks (I. Goodfellow)

Why Generative Models?
● Model very complex and high-dimensional distributions.
● Be able to generate realistic synthetic samples
○ possibly perform data augmentation
○ simulate possible futures for learning algorithms
● Unsupervised learning: fill blanks in the data
● Manipulate real samples with the assistance of the generative model
○ Example: edit pictures with guidance (photoshop super pro level)

Image inpainting
17
Recover lost information/add enhancing details by learning the natural distribution of pixels.
original enhanced

Speech Enhancement
18
Recover lost information/add enhancing details by learning the natural distribution of audio samples.
original
enhanced

Speech Synthesis
19
Generate spontaneously new speech by learning its natural distribution along time.
Speech
synthesizer

Image Generation
20
Generate spontaneously new images by learning their spatial distribution.
Image
generator
Figure credit: I. Goodfellow

Super-resolution
21
Generate high resolution image version, even introducing made up plausible details.
(Ledig et al. 2016)

23
We will see models that learn the probability density function:
● Explicitly (we impose a known loss function)
○ (1) With approximate density → Variational
Auto-Encoders (VAEs)
○ (2) With tractable density & likelihood-based:
■ PixelCNN, Wavenet.
■ Flow-based models: real-NVP, GLOW.
● Implicitly (we “do not know” the loss function)
○ (3) Generative Adversarial Networks (GANs).
Taxonomy

Auto-Encoder Neural Network
Autoencoders:
● Predict at the output the
same input data.
● Do not need labels
cx x^
Encode Decode
25
Bottleneck

● Q: Is an AE a generative model? (How can we make it generate data?)
c
Encode Decode
“Generate” 26

● Q: Is an AE a generative model? (How can we make it generate data?)
○ A: This “just” memorizes codes C from our training samples!
c
Encode Decode
“Generate”
memorization
27

Variational Auto-Encoder
VAE intuitively:
● Introduce a restriction in z, such that our data points x (e.g. images)
are distributed in a latent space (manifold) following a specified
probability density function Z (normally N(0, I)).
z
Encode Decode
z ~ N(0, I)
28

VAE intuitively:
● Introduce a restriction in z, such that our data points x (e.g. images)
are distributed in a latent space (manifold) following a specified
probability density function Z (normally N(0, I)).
z
Encode Decode
z ~ N(0, I)
We can then sample z to
generate NEW x data
points (e.g. images).
29

● VAE, aka. where Bayesian theory and deep learning collide, in depth:
Fixed
parameters
Generate X,
N times
sample z,
N times
Deterministic
function 30Credit:

● VAE, aka. where Bayesian theory and deep learning collide, in depth:
Fixed
parameters
Generate X,
N times
sample z,
N times
Deterministic
function
Maximize the probability of
each X under the generative
process.
Maximum likelihood framework

Intuition behind normally distributed z vectors: any output distribution can be achieved from the simple
N(0, I) with powerful mappings.
N(0, I)
32

Intuition behind normally distributed z vectors: any output distribution can be achieved from the simple
N(0, I) with powerful mappings.
Who’s the strongest non-linear and learnable mapper in the universe (so far)?
33

Now to solve the maximum likelihood problem… We’d like to know and . We
introduce as a key piece → sample values z likely to produce X, not just the whole
possibilities.
But is unkown too! Variational Inference comes in to play its role: approximate
with .
Key Idea behind the variational inference application: find an approximation
function that is good enough to represent the real one → optimization problem.
35

Neural network prespective
The approximated function starts to shape up as a neural encoder, going from training datapoints x to
the likely z points following , which in turn is similar to the real .
36Credit: Altosaar
What is a variational autoncoder? (Altosaar 2017)

Neural network prespective
The (latent→ data) mapping starts to shape up as a neural decoder, where we go from our sampled z
to the reconstruction, which can have a very complex distribution.
37Credit: Altosaar

Continuing with the encoder approximation , we compute the KL divergence with the true
distribution:
38
KL divergence
Credit: Kristiadis

distribution:
Bayes rule
39
Variational autoencoder: Intutition and implementation (Kristiadi 2017)

distribution:
Gets out of expectation for no
dependency over z.
40Credit: Kristiadi

distribution:
41Credit: Kristiadi

distribution:
A bit more rearranging with sign and grouping leads us to a new KL term between
and , thus the encoder approximate distribution and our prior. 42
Credit: Kristiadi

We finally reach the Variational AutoEncoder objective function.
43Credit: Kristiadi

log likelihood of our data
44Credit: Kristiadi

Not computable and non-negative (KL)
approximation error.
45Credit: Kristiadi

46
Reconstruction loss of our data
given latent space → NEURAL
DECODER reconstruction loss!
Credit: Kristiadi

Reconstruction loss of our data
given latent space → NEURAL
DECODER reconstruction loss!
Regularization of our latent
representation → NEURAL
ENCODER projects over prior.
47Credit: Kristiadi

Now, we have to define shape to compute its divergence against the prior (i.e. to properly
condense x samples over the surface of z). Simplest way: distribute over normal distribution with
predicted moments: and .
This allows us to compute the KL-div with in a closed form!
48Credit: Kristiadi

z
Encode Decode
We can compose our encoder - decoder setup, and place our VAE losses to regularize and reconstruct.
49

z
Encode Decode
Reparameterization trick
But WAIT, how can we backprop through sampling of ? Not differentiable!
50

z
Reparameterization trick
Sample and operate with it, multiplying by and summing
51

z
Generative behavior
Q: How can we now generate new samples once the underlying generating distribution is learned?
52

z1
Generative behavior
Q: How can we now generate new samples once the underlying generating distribution is learned?
A: We can sample from our prior, for example, discarding the encoder path.
z2
z3
53

54
Walking around z manifold dimensions gives us spontaneous generation of samples with different
shapes, poses, identitites, lightning, etc..
Examples:
MNIST manifold: https://youtu.be/hgyB8RegAlQ
Face manifold: https://www.youtube.com/watch?v=XNZIN7Jh3Sg

55
Example with MNIST manifold

56
Example with Faces manifold

Code show with PyTorch on VAEs!
https://github.com/pytorch/examples/tree/master/vae
57

58
Loss
Model MNIST:Binary reconstruction case (BCE loss)

GANs are coming in DGMs II ...
The GAN epidemic

References
61
● NIPS 2016 Tutorial: Generative Adversarial Networks (Goodfellow 2016)
● Auto-Encoding Variational Bayes (Kingma & Welling 2013)
● https://wiseodd.github.io/techblog/2016/12/10/variational-autoencoder/
● https://jaan.io/what-is-variational-autoencoder-vae-tutorial/
● Tutorial on Variational Autoencoders (Doersch 2016)

Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018

More Related Content

What's hot (20)

Similar to Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018 (20)

More from Universitat Politècnica de Catalunya (20)

Recently uploaded (20)

Variational Autoencoders VAE - Santiago Pascual - UPC Barcelona 2018