THYNK UNLIMITED

WE LEARN FOR THE FUTURE

AUTOENCODERS,
GANS AND
DIFFUSION MODELS
PRESENTATION
PRESENTED BY:

ABDELRAHMAUN HASSAAN ,
YUNIS KAHALAN, MOHAMED
HAMILA
 PRESENTATION OUTLINE

Mohamed Hamila: Autoencoders

Yunis Kahalan: GANs
Abderlrahman Hassaan: Diffusion Models
WHAT ARE
AUTOENCODERS
Autoencoders are neural networks designed
to learn efficient representations of data,
typically by compressing it into a lower-
dimensional space and then reconstructing
the original data from this compressed
representation.

They are designed to efficiently compress

(encode) and decompress (decode) data by
learning an approximation to the identity function
over the data.
AUTOENCODER COMPONENTS

01 02 03
ENCODER CODE DECODER
goal is to capture the most also known as the latent goal is to to reconstruct the original
important features of the input data representation or latent space, is the input data from the encoded
in the code while discarding compressed representation of the representation (code).
unnecessary details. input data generated by the encoder.
 TYPES OF AUTOENCODERS IN DEEP
LEARNING
Vanilla Autoencoder Sparse Autoencoder

Denoising Autoencoder
 PROS AND CONS
VANILLA

good
reconstruction
quality

simplicity
limited ability
lack or to capture
regulization complex
simplicity patterns
 PROS AND CONS

SPARCE
feature
selection

regulariztion
diifficulty in increased
tuning computational
sparsity complexity
improved parameters
generalization
 PROS AND CONS

DENOISING
robustness to
noise

improved
reconstruction
quality complexity in sesnitivity to
noise noise type and
data cancelling level
augmentation
BUILDING AND
TRAINING
BUILDING AND
TRAINING
REAL LIFE APLICATIONS
IMAGE DENOISING
medical imaging
photography
GANS

INTRODUCTION MATHEMATICAL COMPOMENTS

FORMULATION
What are GANs/Idea Formulation Generator/Discriminat
or

LOSS FUNCTION TRAINING DYNAMICS APPLICATIONS

Predictive Tools Training/Convergence Real world
Maintenance
WHAT IS A GAN
What are GANs?
•Designed by Ian Goodfellow and his colleagues in 2014.

•They consist of two neural networks, termed the

generator and the discriminator.

Core Idea:
•The generator tries to fool discriminator
MATHEMATICAL FORMULATION

•Formulation: GANs are formulated as a minimax game where

the generator G and the discriminator D have competing goals.
•The value function V(D,G) represents this game:
COMPONENTS OF A GAN
•Generator (G):
•Aim: To map latent space vector z to data space
to fool the discriminator.
•Function: G(z;θ) where z is sampled from a noise
distribution p(z), and θ​ are the parameters of G.
•Discriminator (D):

•Aim: To estimate the probability that a given

sample came from the training data rather than G.
•Function: D(x;θ) outputs a scalar representing the
probability that x came from the data rather than p.
LOSS FUNCTION IN GANS
••Loss Functions:
•The discriminator's loss:

•The generator's loss is to maximize what discriminator is

minimizing:
TRAINING DYNAMICS

•Training GANs:
•The training involves alternating between optimizing D to
maximize log⁡D (x) for real x and log⁡( 1−D(G(z))) for generated x
by G, and optimizing G to minimize log⁡( 1−D(G(z))).

•Convergence:
•The training of GANs continues until a point at which the
discriminator is fooled about half the time
APPLICATIONS OF GANS
•Real-world Applications:
Image synthesis and manipulation, style transfer, and photo-
realistic images.

Super-resolution, data augmentation, and generating training

data for other models.

Malware sample generation for training nalware detection

engines
DIFFUSION MODELS

WHAT ARE DIFFUSION HOW DO DIFFUSION

INTRODUCTION MODELS IN MACHINE MODELS WORK?
LEARNING?

BENEFITS OF USING APPLICATIONS OF

KEY TAKEAWAYS
DIFFUSION MODELS DIFFUSION MODELS
WHAT ARE DIFFUSION MODELS IN
MACHINE LEARNING?
Diffusion models are generative models, which means that they
generate new data based on the data they are trained on. For
example, a diffusion model trained on a collection of human faces
can generate new and realistic human faces with various features
and expressions, even if those specific faces were not present in
the original training dataset.
DENOISING DIFFUSION PROBABILISTIC
MODELS (DDPMS)
SCORE-BASED GENERATIVE MODELS
(SGMS)
HOW DO DIFFUSION
MODELS WORK

Data Forward Training the Reverse

Preprocessing Diffusion Model Diffusion
The initial step involves During forward Training a diffusion Once the forward
preprocessing the data to diffusion, the model model involves diffusion process
ensure proper scaling and starts with a sample learning the generates a sample from
centering. Typically, from a simple parameters of the the complex data
standardization is applied to distribution, typically invertible distribution, the reverse
convert the data into a a Gaussian transformations and diffusion process maps it
distribution with a mean of distribution other model back to the simple
zero and a variance of one components. distribution through a
sequence of inverse
transformations.
BENEFITS OF USING
DIFFUSION MODELS

Stable Privacy-Preserving Handling Robustness to

Training Data Generation Missing Data Overfitting
Training diffusion models is Diffusion models are suitable Diffusion models can handle Traditional generative
generally more stable than for applications in which data missing data during the models like GANs can be
training GANs, which are privacy is a concern. Since the generation process. Since prone to overfitting, in
notoriously challenging. GANs model is based on invertible reverse diffusion can work which the model
require balancing the learning transformations, it is possible to with incomplete data samples, memorizes the training
rates of the generator and generate synthetic data the model can generate data and fails to
discriminator networks samples without exposing the coherent samples even when generalize well to unseen
underlying private information parts of the input data are data
of the original data. missing.
APPLICATIONS OF
DIFFUSION MODELS

TEXT TO VIDEO IMAGE TO IMAGE IMAGE SEARCH

Diffusion models offer a
Diffusion models are a Diffusion models are
powerful approach for
promising approach for powerful content-
image-to-image
text-to-video synthesis based image retrieval
translation tasks, which
techniques that can be
involve transforming
applied to image
images from one domain
search tasks. Using the
to another while
reverse diffusion
preserving semantic
process
information and visual
coherence
WELL-KNOWN DIFFUSION
MODELS FOR IMAGE
GENERATION
Stable Diffusion DALL-E 2 Imagen GLIDE
Stable diffusion is a DALL-E 2 utilizes Imagen is a text-to-image Guided Language to
popular approach for contrastive models diffusion model that Image Diffusion for
image generation that like CLIP to learn stands out for its Generation and
uses diffusion models robust image exceptional image Editing (GLIDE) is
(DMs) and the efficiency representations that generation capabilities. another powerful
of latent space capture semantics The model is built upon text-conditional
representation and style two key components image synthesis
model by OpenAI.
STABLE DIFFUSION
DALL-E 2
IMAGEN
GLIDE
DIFFUSION MODELS: KEY
TAKEAWAYS
Diffusion models are a class of generative
models that simulate the data generation
process by transforming a simple starting
distribution into the desired complex data
distribution through a sequence of
invertible operations.
Compared to traditional generative
models, diffusion models have better
image quality, interpretable latent space,
and robustness to overfitting.
Thank you Any
Questions