Well come to my presentation Realizing a

communication system with deep learning

1
By Abu sheka

Advisor: Dr.Sultan Feiso

11/02/2023
 “ Acknowledgment
 First and foremost, I would like to thank God. He has given me
strength and encouragement throughout all the challenging moments

2
and endless love, mercy, and grace
”
of completing this thesis Iam truly grateful for His unconditional

 I would like to express my sincere gratitude to my advisor, Sultan

Feiso,(PhD),for his unwavering support and guidance throughout
my research journey.

 Most of all, this thesis is dedicated to my beloved family

11/02/2023
 01
Abstract
3

02 Backgrounds

Literature Review
contents 03

04 Methodology

05
Result and conclusion
11/02/2023
 Abstract
4
 we create a complete wireless communication system using deep neural networks (DNNs) to handle essential tasks like
encoding, decoding, modulation, and demodulation.
 Yet, accurate channel state information (CSI) is required for optimal decoding, posing challenges due to its temporal and
spatial variation.
 To address this, we suggest using a conditional generative adversarial network (GAN) to simulate channel effects.
 This GAN connects transmitter and receiver DNNs, enabling gradient back-propagation
 A conditional GAN models channel effects, incorporating pilot symbol-based received signals.
 Convolutional layers handle high dimensionality for long transmit sequences
 Our approach performs well on different channel types, including additive white Gaussian noise (AWGN) channels,
Rayleigh fading channels, and frequency-selective channels
 By implementing the CNN algorithm in Python,
 the proposed aproach shows good performance when compared to conventional communication system
 This introduces new possibilities for data-driven DNNs in end-to-end communication systems.

11/02/2023
 Background
5
 Deep learning a fresh paradigm for upcoming communication systems in addition
to improving the conventional communication building blocks.
 By optimizing an end-to-end loss function, a deep learning model's features and
parameters can be learned entirely from the data without the use of manual labor or
not planned designs
 In several industries, particularly communication systems, deep learning has
emerged as a game-changing technology.
 Generative Adversarial Networks (GANs) are a powerful class of neural networks
that are used for unsupervised learning. It was developed and introduced by Ian J.
Good fellow in 2014.

11/02/2023
 Statement of the problem
6

 One of the problem in the end-to-end communication

system is the issue of high dimensionality that arises during
training when the transmitted symbol sequence is extensive.
 The second challenge in modulation classification has been
dealing with noisy signals .
 However, the introduction of a convolutional neural network
(CNN) model has significantly addressed this issue

11/02/2023
 Objective of the study
7
 The General objective of my thesis is to develop and implement a
communication system that uses deep learning techniques and realizing
it.
 Specific objectives :
 To apply deep learning techniques to a communication system consisting
of a whole transmitter and receiver.
 To compare the performance of the conditional GAN-based
communication system with other existing communication systems to
showcase its advantages.
 To evaluate the performance of the end-to-end communication system by
measuring its ability to handle different channel conditions .11/02/2023
8 Significance of the Study
 Since deep learning is an emerging technology especially in
communication system it’s importance is very crucial.
 To enhance communication performance, adaptability, reliability,
addressing the limitations of conventional methods.
 deep learning with a communication system has the potential to
transform the way we design, implement, and operate communication
systems,
 improving their performance, efficiency, and security, and opening up
new possibilities for innovative applications.

11/02/2023
9 Scope of the study
 Deep learning models typically require a large amount of labeled data for
training.
 When dealing with communication systems, unexpected conditions or
new communication protocols may cause the model to perform poorly.
 especially if the training data is limited in diversity.

 Deep learning models, particularly complex.

 Meaning can be challenging to understand how they arrive at specific

decisions 11/02/2023
10 Contribution of the thesis
 Introduces the use of conditional generative adversarial networks
(GAN) to model the conditional distribution, p(y|x), of wireless
channels.
 In this thesis we develops an end-to-end learning-based
communication system, leveraging the conditional GAN's modeling
of the channel's conditional distribution.
 Convolutional neural networks (CNNs) are employed to address the
curse of dimensionality.
11/02/2023
 11 LITERATURE REVIEW
 in the author Oshea et al. [1] proposed joint transmitter-receiver optimization

via a neural network, achieving near-optimal performance in diverse tasks.

 Zhang et al. exhibited deep learning's potential in improving underwater

acoustic communication's challenging conditions.

 while these results are promising, challenges persist.

 This cause model interpretability remains complex, making it challenging to

understand network decisions.

11/02/2023
 End-to-end learning-based communication systems
12

 As read from author (I. Goodfellow et al) introduction of Generative

Adversarial Networks (GANs) in marked a pivotal development in generative
models.
 two neural networks ;Generator and discriminator;

 The generator's role is to craft synthetic data samples, like images, mirroring
real data from a designated dataset.
 While , the discriminator learns to differentiate between genuine data samples
11/02/2023

and synthetic ones produced by the generator.

13 Learning-Based Decoders(LBD)
 Learning-based encoding and decoding approaches, which have been utilized to
improve decoding performance in communication systems.
 dating back to the 1990s, with attempts to use Recurrent Neural Networks
(RNNs) for decoding codes.
 In recent years, deep learning approaches, particularly Deep Neural Networks
(DNNs), have been widely employed in decoding with various applications in
coding theory.
 Here are some examples of learning-based decoding techniques and their
11/02/2023

applications:
14
Cont….
 However, decoding long codewords using standard neural networks faces
the curse of dimensionality.
 To address this, a partition method is employed to train longer
codewords.
 RNNs are used for decoding convolutional and turbo codes.

 Recurrent neural networks can effectively capture temporal

dependencies, .
11/02/2023
15 Cont…
 Our approach has two key improvements that address real-world wireless

communication challenges and broaden the scope of learning-based encoding

and decoding.

 1.Convolutional Layers :This is crucial for achieving higher data rates and

better error correction in practical communication systems.

 2.Extension to Frequency-Selective Channels with Inter-Symbol Interference

(ISI):This extension equips our method to effectively tackle complex channel

conditions and improve overall performance. 11/02/2023
16 Methodology

11/02/2023
17 MODELING CHANNEL WITH CONDITIONAL
GAN DNNs that are optimized for the transmitter and
 develops
receiver.
We utilize a conditional GAN to learn the channel effects
 By using the conditional GAN, it is possible to learn the
channel's output distribution in a data-driven way
 This address numerous complex channel effects.

11/02/2023
18 Cont…
 the GAN system consists of two DNNs, i.e., the generator, G, and the
discriminator, D.
 The input to the generator, G, is a noise vector z sampled from a prior
distribution , e.g., uniform distribution
 The input z is transformed by the generator into a generated sample,
G(z).
 During the training, the discriminator, D, learns to distinguish the data
generated by the generator and the data from the target distribution
 while the generator, G, learns to generate samples to fool the
discriminator into making mistakes.
11/02/2023
 Cont…
19
 the objective functions for training the generator and the discriminator
can be expressed as
 LG = Ez∼pz [log(1 − D(G(z)))],
 LD = Ex∼pdata [log(D(x))] + Ez∼pz [log(1 − D(G(z)))],
 The objective of the discriminator, D, is to give a high value when the
input belongs to the real dataset
 while the objective of generator, G, is to maximize the output of the
discriminator, D, given the generated samples, G(z).

11/02/2023
20
Cont…

 Therefore, the output of the G will be G(x,m) and the output of D will
be D(x,m). The optimization objective functions for the generator and
the discriminator become
 LG = Ez∼pz [log(1 − D(G(z,m) , m))]

 LD = Ex∼pdata [log(D(x,m))] + Ez∼pz [log(1 − D(G(z,m) ,m))],

11/02/2023
21
Convolutional Layers Based Channel GAN
 Here layers have been introduced to efficiently extract features for
images based on their shared-weight
 In a fully connected layer, each neuron is connected to all neurons in
the previous layer
 u(i)[n] =( n-k), |

11/02/2023
22 Training Receiver

 At the receiver, a DNN model is trained for recovering the transmitted

signal s
 where the input is the received signals corresponding to the transmitted
data, y, while the output is the estimation sˆ.
 The receiver can be trained easily since the loss function is computed at
the receiver and thus the gradients of the loss can be easily obtained.

11/02/2023
 23 Training Transmitter

 trained similarly to the receiver using the channel generator as a substitute

channel
 The transmitter, generator, and receiver can all be seen as one DNN during
training
 The values of the final hidden layer in the transmitter's DNN serve as its
output.
 The gradients are transmitted back to the transmitter via the conditional
GAN, 11/02/2023
24 Training Channel GAN
 the channel GAN training process is thoroughly described.
 The generator and discriminator are trained iteratively during each
cycle.
 While training the other model, the parameters of one model will be
fixed.
 the learnt transmitter, the encoded signal from the transmitter traveling
via the actual channel can be used to retrieve the true data.
 while the encoded signal traveling through the channel generator is
used to obtain the fake information

11/02/2023
25 Results and Discussion
 Here two different types of DNN models are created.
 Convolutional neural network (CNN) and Fully connected Network(FCN).
 To escape the curse of dimensionality the FCN is utilized for small block
sizes and the CNN for high block sizes.
 Adam updates the weights of both models, and 320 batches are used for
training.
 In our tests, three different channel types
 AWGN channels,
 Rayleigh channels, and
 frequency-selective multipath channels, 11/02/2023
26 Baselines
 The end-to-end learning-based communication system is compared to
the conventional communication system,
 which is made up of various signal processing modules, each of
which is created based on the channel's prior knowledge.
 Under each type of channel, the bit-error rate (BER) and block-error
rate (BLER) are contrasted.
 In our basic system, the Hamming code or convolutional codes are
utilized with 4-QAM as the modulation.
 The Viterbi method is used for maximum likelihood sequence
decoding for convolutional codes.
11/02/2023
27
Cont…

 Channel coding makes use of the rate-1/2 recursive systematic

convolutional (RSC) code.
 The ISI in the frequency-selective multipath channel is addressed using
OFDM.
 The received signal can be written as = h + ,

 The auto-encoder system is trained for the AWGN channel by directly

introducing Gaussian noise into the hidden layer
11/02/2023
28 End-to-End Communication System
 Three different types of channels

 AWGN channel,

 the Rayleigh fading channel, and

 the frequency-selective multipath channel

 are used to construct a channel-agnostic end-to-end communication

system based on the channel GAN

11/02/2023
29  AWGN Channel

11/02/2023
30  First we use FCN for small block size.
The BLER and BER of learning based approach is
similar to Hamming code (7,4) MLD.
This proves the effectiveness of the surrogate.
channel

11/02/2023
31 Rayleigh Fading Channel
 Here employing CNNs for transmission with a larger block size and incorporating channel encoding
 A comparison is made between the baseline method using RSC with a code rate of 1/2 and QAM.

11/02/2023
32 frequency-selective multipath.
 using Orthogonal Frequency Division Multiplexing (OFDM) as the
baseline system.
 the OFDM system employs 4-QAM modulation and features 64
subcarriers, each with a cyclic prefix duration of 16.
 For channel coding, Recursive Systematic Convolutional (RSC) codes
are employed.
 Results underscore the potential of end-to-end communication systems
to deliver enhanced performance in frequency-selective fading
channels
11/02/2023
33

11/02/2023
34 Conclusion
 Our strategy employs conditional GANs to emulate channel distributions
and generate data tailored to instant channel states through pilot
information.
 encompasses transmitter, channel GAN, and receiver DNNs, honed
iteratively to minimize end-to-end loss.
 The simulation results spanning various channels AWGN, Rayleigh fading,
and frequency-selectivtive
 Shows superior performance compared to conventional methods

11/02/2023
35 Future Directions and Recommendations
The end-to-end pipeline consists of DNNs for the transmitter, the channel GAN,
and the receive.
 By iteratively training these networks, the end-to-end loss can be optimized
in a supervised way.
 The simulation results on the AWGN channels, Rayleigh fading channels, and
frequency-selective channels confirm the effectiveness of the proposed
method,
 by showing similar or better performance compared with the traditional
approaches, which are designed based on expert knowledge and channel
models
 the future is to validate the proposed method by testing it with actual data
derived from real-world communication scenarios. 11/02/2023
Thank You!!