MIMO in 5G: Revolutionising Wireless Communication

A TECHNICAL REPORT
submitted by
Kapil Kumar
Mahto(AM.EN.U4ECE21126), Midhin
Manoj(AM.EN.U4ECE21134), Anandhu S
(AM.EN.U4ECE21107), Reshma S
Kumar(AM.EN.U4ECE21148) under the
guidance of

Aswathy K Nair

submitted as part of
19ECE495 PROJECT PHASE I
in
ELECTRONICS AND COMMUNICATION ENGINEERING

AMRITA SCHOOL OF ENGINEERING

AMRITA VISHWA VIDYAPEETHAM
AMRITAPURI (INDIA)
JULY– 2024
 AMRITA SCHOOL OF ENGINEERING

AMRITA VISHWA VIDYAPEETHAM

AMRITAPURI (INDIA)

BONAFIDE CERTIFICATE
This is to certify that the report entitled “MIMO in 5G: Revolutionising
Wireless Communication” submitted by Kapil Kumar
Mahto(AM.EN.U4ECE21126), Midhin Manoj(AM.EN.U4ECE21134), Anandhu
S (AM.EN.U4ECE21107), Reshma
S Kumar(AM.EN.U4ECE21148) as part of the 19ECE495 PROJECT PHASE I is
a bonaLide record of the work carried out by her under my guidance
and supervision at Amrita School of Engineering, Amritapuri.

Signature of Supervisor:
Guides name: Prof. Aswathy
K Nair Designation:
Assistant Professor
Department of ECE

Signature of Co-Supervisor:
Name of co supervisor: Aswathy K Nair , Baby Sreeja
Designation: Assistant Professor
Department of
ECE Date:
 AMRITA SCHOOL OF ENGINEERING

AMRITA VISHWA VIDYAPEETHAM

AMRITAPURI - 690 542

DEPARTMENT OF ECE

DECLARATION

We, Kapil Kumar Mahto(AM.EN.U4ECE21126), Midhin Manoj(AM.EN.U4ECE21134),

Anandhu S (AM.EN.U4ECE21107), Reshma S Kumar(AM.EN.U4ECE21148), hereby
declare that this technical report entitled “MIMO in 5G: Revolutionising
Wireless Communication Report”, is the record of the original work done by
us under the guidance of Aswathy K Nair, Assistant Professor, Department
of ECE, Amrita School of Engineering, Amritapuri. To the best of my
knowledge this work has not formed the basis for the award of any
degree/diploma/ associateship/fellowship/or a similar award to any
candidate in any University.

Place: Signature of the Students

Date:
ABSTRACT
The advent of 5G technology marks a significant leap forward in wireless

communication, promising unprecedented data rates, reduced latency, and enhanced

connectivity. A key enabler of these advancements is Multiple Input Multiple Output

(MIMO) technology, which leverages multiple antennas at both the transmitter and

receiver ends to improve communication performance. This paper explores the

transformative impact of MIMO on 5G networks, highlighting its role in enhancing

spectral efficiency, reliability, and network capacity. By exploiting spatial diversity and

multiplexing, MIMO enables the simultaneous transmission of multiple data

streams, thus overcoming the limitations of previous generations. This study

delves into the technical underpinnings of MIMO, its integration into the 5G

architecture, and the resulting benefits in various application scenarios. Furthermore, it

addresses the challenges associated with MIMO implementation, including hardware

complexity, signal processing demands, and interference management. As 5G

networks continue to roll out globally, MIMO stands as a cornerstone technology,

driving the evolution of mobile communication towards a more connected and

intelligent future.
 Contents

Acknowledgement

1. Abstract
1.1 Introduction to report..........................................................................1

1.2 Basic Terminologies..............................................................................1

2. Literature Survey

2.1.1 Paper 1: User-Centric Cell-Free Massive MIMO Networks: A Survey of Opportunities,

Challenges and Solutions............................................................... 1

2.1.2........................................................................................................Paper 1 Abstract 1

2.1.3........................................................................................................Paper 1
INTRODUCTION....................................................................................................1

2.1.4........................................................................................................Paper 1
METHODOLOGY...................................................................................................1

2.1.5.........................................................................................................Paper 1 RESULTS
AND DISCUSSION....................................................................................................2

2.1.6.........................................................................................................Paper 1 CONCLUSION
AND FUTURE WORK................................................................................................2

2.2.1.................................................................................................................. P

aper 2: A Review of the State of the Art and Future Challenges of Deep

Learning-Based Beamforming................................................................................................ 2

2.2.2........................................................................................................Paper 2 Abstract 2

2.2.3................................................................................................................. Paper 2
INTRODUCTION...................................................................................................2

2.2.4 Paper 2 STATE OF THE ART IN DEEP LEARNING-BASED BEAMFORMING

.........................................................................................2

2.2.5 Paper 2 FUTURE CHALLENGES......................................................................................... 2

2.2.6 Paper 2 CONCLUSION................................................................................................3

2.3.1 Paper 3:3D Beamforming Technologies and Field Trials

in 5G Massive MIMO Systems..............................................................................................3
2.3.2 Paper 3 Abstract. . . . . . . . . . . . . . . . . . .. . . . .. . . . . . . . .............................................3

2.3.3 Paper 3 INTRODUCTION...........................................................................................3

2.3.4 Paper 3 METHODOLOGY..........................................................................................3

2.3.5 Paper 3 RESULTS......................................................................................................3

2.3.6 Paper 3 CONCLUSION AND FUTURE WORK...............................................................................................3

2.4.1 Paper 4: DNN-SNN CO-LEARNING FOR SUSTAINABLE SYMBOL DETECTION IN 5G SYSTEMS

ON LOIHI CHIP.............................................................................................................................. 4

2.4.2 Paper 4 Abstract.......................................................................................................... 4

2.4.3 Paper 4 INTRODUCTION..........................................................................................4

2.4.4 Paper 4 METHODOLOGY.........................................................................................4

2.4.5 Paper 4 RESULTS.....................................................................................................5

2.4.6 Paper 4 CONCLUSION AND FUTURE WORK........................................................................5

3. ACKNOWLEDGEMENT

4. REFERENCES
 ACKNOWLEDGEMENT

First and foremost, we express our deepest gratitude to Amma, for her unwavering support, love, and

encouragement throughout this project.

We extend our heartfelt thanks to the Head of the Department, Dr Ramesh Bhakthavatchalu, for their

invaluable guidance and constant support. We are also deeply grateful to the Vice Head of the

Department, Remya mam, for their insightful suggestions and encouragement.

We sincerely acknowledge the invaluable support and guidance of our project guide, Aswathy K

Nair. Your expertise and advice were instrumental in the successful completion of this project.

We would also like to thank class advisors Prof. Aswathy K Nair and Prof. Baby Sreeja, Department

of Electronics and Communication, for their assistance and encouragement throughout this project.

Our gratitude extends to all the faculty members and the Department of Electronics and

Communication for providing us with the necessary resources and a conducive environment for this

project.
 1

MIMO in 5G: Revolutionising Wireless Communication Report

* This report provides an overview on the current state and advances in MIMO in 5G .

Kapil Kumar Mahto(AM.EN.U4ECE21126), Midhin Manoj(AM.EN.U4ECE21134), Anandu

S (AM.EN.U4ECE21107), Reshma S Kumar(AM.EN.U4ECE21148)
Under the Guidance of Aswathy K Nair , Assistant Professor
Department of Electronics And Communication Amrita Vishwa Vidyapeetham

Abstract— The advent of 5G technology marks a significant

leap forward in wireless communication, promising II. METHODOLOGY
unprecedented data rates, reduced latency, and enhanced The methodology of this paper revolves around a comprehensive
connectivity. A key enabler of these advancements is Multiple examination of various techniques and frameworks employed to
Input Multiple Output (MIMO) technology, which leverages optimise the deployment of Massive MIMO networks. Distributed
multiple antennas at both the transmitter and receiver ends to
resource allocation is a key focus, with game theoretic approaches,
improve communication performance. This paper explores the
such as General Game Theory, Auction Theory, and Stable Matching
transformative impact of MIMO on 5G networks, highlighting
Theory, being highlighted for their ability to model interactions between
its role in enhancing spectral efficiency, reliability, and
users and predict their strategies. However, the paper also
network capacity. By exploiting spatial diversity and
acknowledges the limitations of these approaches, such as the need
multiplexing, MIMO enables the simultaneous transmission of
for feedback channels and convergence towards a Nash equilibrium.
multiple data streams, thus overcoming the limitations of
Non-game theoretic methods, like Decomposition Theory, are used to
previous generations..
Index Terms—MIMO technology's role in 5G, its benefits, address complex problems by breaking them into more manageable
and the challenges faced in its implementation. sub-problems. For instance, the uplink max-min SINR problem is
tackled by separating it into receiver filter design and power allocation
PAPER 1: User-Centric Cell-Free Massive MIMO sub-problems, which are solved using generalised eigenvalue problems
Networks: A Survey of Opportunities, Challenges and and geometric programming methods, respectively.Additionally, local
Solutions beamforming techniques, such as Partial Zero-Forcing (PZF) and
Abstract—The advent of Multiple-Input Multiple-Output Protective Partial Zero-Forcing (PPZF), are analysed for their
(MIMO) technology has marked a significant evolution in effectiveness in mitigating interference, showing comparable
wireless communication systems, playing a pivotal role in the performance to regularised zero-forcing and superior performance to
development of fifth-generation (5G) networks. MIMO other traditional methods. Finally, the paper addresses the challenge
technology enables the simultaneous transmission and of maximising energy efficiency by dividing the problem into sub-
reception of multiple data streams, enhancing spectral problems focused on rate maximisation, energy efficiency without rate
efficiency, data rates, and overall network capacity. In 5G
constraints, and power minimisation, each solved through efficient
networks, massive MIMO takes this concept further by utilising
a large number of antennas at the base stations, which proposed methods.
facilitates advanced beamforming techniques and improved
spatial multiplexing.

I. INTRODUCTION
A user-centric cell-free network architecture defines a cooperative III. RESULTS AND DISCUSSION
serving cluster of spatially distributed transmitters specifically for each
The results obtained from the various studies and simulations
user, hence the name user-centric; this in turn, eliminates conventional conducted reveal significant insights into the performance and
cell-edges, and is, hence, termed cell-free. Based on this definition, each optimisation of massive MIMO networks. For instance, an analysis
user in the network is served by its neighboring set of transmitters, so of antenna density showed that co-located deployment of
each user is found at the effective center of its serving cluster. The user- antennas provides the highest profile of channel hardening,
centric cell-free scheme can be seen as a practical method to deploy cell- followed by distributed deployment with correlated shadowing, and
free communications, and the concept is often coupled with Multiple- finally distributed deployment with independent shadowing.
input multiple-output (MIMO) network concepts, and recently, massive Additionally, the performance improvements observed with lower
MIMO concepts thus branding the architecture as cell-free massive path loss exponent (PLE) indicate better channel hardening for
MIMO. The growing interest in this technology relates to seeing it as one distributed deployments, consistent with prior research
of the drivers for future mobile networks; this interest motivates us to findings.Pilot assignment (PA) policies play a crucial role in
review the developments, current trends, and challenges in this managing co-pilot user interference in cell-free networks. Well-
topic.MIMO technology has been one of the most important structured PA policies have been shown to reduce the required
advancements in wireless communications. It provides the capability for pilot sequence length significantly while achieving negligible
contamination. This is particularly important in scenarios where the
a transmitter to send concurrent signals using the installed multiple
number of users exceeds the available resources, necessitating
antennas, thus allowing for the vectorisation of the transmission. MIMO
efficient user scheduling and pilot assignment strategies.
provides beamforming capabilities where the power of the transmit
signals can be focused toward specific users.
 2

IV. CONCLUSION AND FUTURE WORK

The survey provides a comprehensive overview of the current very difficult to consistently treat changing environments,
state and future directions of user-centric cell-free massive MIMO where the emitters move and therefore the weights must
networks. Key findings highlight that while significant continuously be computed. Not to mention the significant risk
advancements have been made in the theoretical foundations and of performance degradation due to mismatches between the
practical implementations, several challenges persist, particularly
expected and the actual signal steering vectors
in scalability, resource allocation, and integration of advanced
technologies such as machine learning and IRS (Intelligent
Reflecting Surfaces). Scalability issues can be addressed by II. STATE OF THE ART IN DEEP LEARNING-
developing distributed resource allocation schemes that are both BASED BEAMFORMING
efficient and effective in dense deployment scenarios. The role of
machine learning in enhancing various aspects of cell-free 1. Adaptive Beamforming and DOA Estimation
networks is promising but requires further exploration to quantify Traditional adaptive beamforming algorithms and Direction of
performance gains accurately. Additionally, the deployment of IRS Arrival (DOA)estimation methods have been enhanced using
can potentially boost energy efficiency by improving channel ML and DL approaches. These include the use of neural
conditions at a lower operational cost.Future research should networks (NNs) to improve the accuracy and
focus on a few critical areas: the development of comprehensive efficiency of beamforming processes.Significant progress
models for distributed SDN frameworks to dynamically manage has been made in integrating neural network-based and
DU assignments; the exploration of non-uniform quantisation clustering-based approaches for robust mask estimation,
techniques to improve signal compression on the fronthaul; and which is crucial for adaptive acoustic beamforming.
the investigation of high-mobility user performance to address 2. Applications in Various Domains
issues such as channel aging and pilot design. Acoustic Beamforming: Neural networks have been employed
to estimate spectral masks, improving the robustness and
accuracy of acoustic beamforming. For instance, DNN-based
time-frequency mask estimation has shown to be effective in
PAPER 2: A Review of the State of the Art and enhancing speech recognition systems.
Future Challenges of Deep Learning-Based Medical Ultrasound: Deep learning techniques have been
applied to enhance beamforming in medical ultrasound,
Beamforming improving imaging quality and diagnostic capabilities.
Abstract—The key objective of this paper is to explore the Adaptive ultrasound beamforming using deep learning has
recent state-of-the-art artificial intelligence (AI) applications on the demonstrated significant improvements in medical imaging.
broad field of beamforming. Hence, a multitude of AI-oriented Massive MIMO Systems: In 5G communication systems, DL-
beam forming studies are thoroughly investigated in order to based beamforming is essential for managing the high
correctly comprehend and profitably interpret the AI contribution in
propagation losses in the millimeter-wave band. The
complexity of neural networks increases with the number of
the beamforming performance. Starting from a brief overview of
antenna elements, posing challenges in fast training and
beamforming, including adaptive beamforming algorithms and
implementation in dynamic environments.
direction of arrival (DOA) estimation methods, our analysis probes
3. Deep Learning Techniques and Architectures
further into the main machine learning (ML) classes, the basic Various deep learning architectures, including Convolutional
neural network (NN) topologies, and the most efficient deep Neural Networks(CNNs) and Recurrent Neural Networks
learning (DL) schemes. Subsequently, and based on the prior (RNNs), have been explored for their potential in beamforming
aspects, the paper explores several concepts regarding the applications. These architectures can effectively handle the
optimal use of ML and NNs either as standalone beamforming and spatial and temporal characteristics of signals, making them
DOA estimation techniques or in combination with other suitable for real-time beamforming tasks.
implementations, such as ultrasound imaging, massive multiple- Deep Long Short-Term Memory (LSTM) networks have been
input multiple-output structures, and intelligent reflecting surfaces. utilised for adaptive beamforming in multichannel speech
Finally, particular attention is drawn on the realization of recognition, demonstrating the ability to maintain performance
beamforming or DOA estimation setups via DL topologies. The in challenging conditions.
survey closes with various important conclusions along with an
interesting discussion on potential future aspects and promising
research challenges III. FUTURE CHALLENGES
1. Scalability and Complexity:
I. INTRODUCTION As the number of antenna elements in massive MIMO systems
With today’s rapid technological advancements and the increases, the complexity of neural networks required for
tremendous increase in data volume, the main challenge is to beamforming also grows. Developing scalable and efficient
minimize interference and optimize the capacity of wireless neural network architectures that can handle large-scale antenna
communication systems. Also, the demand for transmission arrays remains a significant challenge.
with the highest possible quality and coverage is gradually 2. Real-Time Implementation:
escalating, thus opting for more efficient antennas in order to Achieving real-time performance with deep learning-based
meet all 5G challenges. Lately, beamforming – defined as a beamforming systems requires optimizing the training and
real-time procedure which can create a main lobe that inference processes. This involves reducing the computational
corresponds to the direction of the desired signal and several complexity and latency associated with deep learning models.
nulls toward the directions of interference signals – has gained 3. Robustness and Adaptability:
remarkable recognition in the area of modern wireless Ensuring that deep learning-based beamforming systems can
communications and radar systems. In essence, beamforming adapt to dynamic and fast-changing environments is crucial.
has been comprehensively studied during the last decades by This includes developing models that can quickly re-train or
means of deterministic and evolutionary methods. Despite adjust to new conditions without significant performance
their advantages in calculating optimum weights and degradation.
estimating the direction of arrival (DOA), such techniques are
 4. Integration with Emerging Technologies:
Integrating deep learning-based beamforming with emerging II. METHODOLOGY
technologies such as intelligent reflecting surfaces and hybrid
precoding for millimeter-wave communications is an area of Our methodology encompasses a comprehensive experimental
ongoing research. These integrations aim to further enhance the setup designed to rigorously evaluate the performance of 3D
efficiency and effectiveness of wireless communication beamforming technologies. The hardware configuration
includes state-of-the-art 5G base stations equipped with
massive MIMO arrays, capable of real-time beamforming
adjustments. We conducted field trials in diverse urban and
IV. CONCLUSION suburban environments to capture a wide range of deployment
The integration of deep learning techniques in beamforming has
scenarios. Key performance metrics such as signal-to-noise
shown promising results across various applications, from acoustic ratio (SNR), data throughput, and user coverage were
signal processing to medical imaging and wireless systematically measured. The software used for data collection
communications. However, challenges related to scalability, and analysis was calibrated to ensure accuracy and
real-time implementation, robustness, and integration with new repeatability. This robust methodological framework enables
technologies must be addressed to fully realize the potential of us to draw meaningful conclusions about the real-world
deep learning-based beamforming. Continued research applicability of 3D beamforming.
and development in this field will be essential to overcoming
these challenges and advancing the state-of-the-art in
beamforming technology. III. Results and Discussion
The results from our field trials demonstrate substantial
performance improvements attributable to 3D beamforming.
PAPER 3:3D Beamforming Technologies and Field Trials Compared to traditional 2D beamforming methods, 3D
in 5G Massive MIMO Systems beamforming exhibited a significant increase in SNR, leading
Abstract—The transition to 5G technology heralds a new era to enhanced data throughput and more reliable user coverage.
of wireless communication, characterized by the need for Specifically, we observed an average throughput gain of 30%
higher data rates and enhanced connectivity. This paper in dense urban areas and 20% in suburban settings. These
explores the implementation and efficacy of 3D beamforming findings underscore the advantages of incorporating the
technologies in massive Multiple Input Multiple Output (MIMO) elevation dimension in beam steering, particularly in complex
systems, a pivotal advancement for 5G networks. By directing environments with varied user distributions. Our discussion
beams in both azimuth and elevation planes, 3D beamforming highlights the practical implications of these results,
significantly mitigates interference and optimizes signal suggesting that 3D beamforming can effectively address the
strength for users. Through extensive field trials in varied challenges of high-density user scenarios and interference
urban and suburban environments, we demonstrate that 3D mitigation. Moreover, we identify certain limitations related to
beamforming achieves substantial improvements in signal-to- the computational complexity and real-time processing
noise ratio (SNR), data throughput, and user coverage. Our requirements, proposing avenues for future optimization.
findings indicate an average throughput increase of 30% in
urban areas and 20% in suburban settings compared to
traditional 2D beamforming methods. Despite the IV. CONCLUSION AND FUTURE WORK
computational complexities associated with real-time
processing, the benefits of 3D beamforming are evident, In conclusion, our research substantiates the critical role of
providing a scalable solution to meet the burgeoning demands 3D beamforming technologies in the evolution of 5G massive
of 5G communication. This paper concludes by proposing MIMO systems. The empirical evidence from our field trials
future research directions, including advanced algorithm confirms that 3D beamforming not only enhances network
development and machine learning integration, to further performance but also provides a scalable solution to meet the
enhance the performance and applicability of 3D beamforming increasing demands for high-speed, reliable communication.
By leveraging the additional degree of freedom in beam
in 5G networks..
steering, network operators can achieve superior signal
quality and user experience.
I. INTRODUCTION
The advent of 5G technology promises significant Looking ahead, further research is essential to address the
advancements in wireless communication, necessitating computational challenges associated with real-time 3D
innovative solutions to meet the growing demand for higher beamforming. Future work will focus on developing advanced
data rates and improved connectivity. Among these algorithms that optimize beamforming performance while
minimizing processing delays. Additionally, extensive field
innovations, 3D beamforming stands out as a critical technique
trials in diverse geographic locations will be conducted to
in massive Multiple Input Multiple Output (MIMO) systems,
validate the scalability and robustness of 3D beamforming in
capable of enhancing network capacity and coverage. By different deployment scenarios. Integration of machine
dynamically steering the transmission and reception beams in learning techniques for adaptive beam management and the
both azimuth and elevation planes, 3D beamforming reduces exploration of hybrid beamforming approaches are also
interference and optimizes signal strength for individual users. promising directions for future investigation. These efforts will
This paper investigates the efficacy of 3D beamforming pave the way for the commercial deployment of 3D
technologies through extensive field trials, offering insights into beamforming, ultimately contributing to the realization of the
their potential to revolutionize 5G networks. full potential of 5G networks.
 PAPER 4:DNN-SNN CO-LEARNING FOR II. METHODOLOGY
SUSTAINABLE SYMBOL DETECTION IN 5G SYSTEMS In recently DNNs have achieved remarkable success in
ON LOIHI CHIP various applications. DNN-based algorithms for signal
Abstract—Performing symbol detection for multiple-input and processing in OFDM-MIMO systems have gained immense
multiple-output orthogonal frequency division multiplexing (MIMO- attention due to their high performance, which is heavily
OFDM) systems is challenging and resource-consuming. In this dependent on the number of neurons and layers in the
paper, we present a liquid state machine (LSM), a type of reservoir network. Consequently, the computational and space
computing based on spiking neural networks (SNNs), to achieve complexity of a high-performance DNN model is very high,
energy-efficient and sustainable symbol detection on the Loihi chip making them impractical for real-time inference with high
for MIMO-OFDM systems. SNNs are more biologically plausible energy efficiency in typical OFDM-MIMO systems that lack
sufficient computation and storage resources.
and energy-efficient than conventional deep neural networks
SNNs, which mimic the neuronal dynamics of the brain, offer a
(DNNs) but have lower performance in terms of accuracy. To
promising alternative. Efficient hardware implementation of
enhance the accuracy of SNNs, we propose a knowledge SNNs has been demonstrated, showcasing their higher
distillation training algorithm called DNN-SNN co-learning, which biological plausibility and energy efficiency compared to
employs a bi-directional learning path between a DNN and an SNN. DNNs. LSMs, a type of RC using SNNs, capture both
Specifically, the knowledge from the output and intermediate layer temporal and spatial information from input data, making them
of the DNN is transferred to the SNN, and we exploit a decoder to useful for restoring corrupted symbols in OFDM-MIMO
convert the spikes in the intermediate layers of an SNN into real systems.
numbers to enable communication between the DNN and the SNN. However, SNNs face training challenges due to the non-
Through the bi-directional learning path, the SNN can mimic the differentiable nature of spike activities, often resulting in lower
behavior of the DNN by learning the knowledge from the DNN. precision compared to DNNs. While the STDP learning
Conversely, the DNN can better adapt itself to the SNN by using algorithm is popular for its simplicity, it alone is insufficient for
the knowledge from the SNN. We introduce a new loss function to high-performance SNN models. Surrogate gradient descent
enable knowledge distillation on regression tasks. Our LSM is training algorithms have been proposed to address this by
implemented on Intel’s Loihi neuromorphic chip, a specialized approximating spiking activities, allowing the back-propagation
hardware platform for SNN models. The experimental results on through time algorithm to train SNN models.
symbol detection in MIMO-OFDM systems demonstrate that our To bridge the performance gap between DNNs and SNNs, we
LSM on the Loihi chip is more precise than conventional symbol propose a KD algorithm called DNN-SNN co-learning. This
detection algorithms. Also, the model consumes approximately six algorithm builds a bi-directional learning path between a DNN
(teacher model) and an SNN (student model). The DNN-SNN
times less energy per sample than other quantized DNN-based
models with comparable accuracy. co-learning algorithm involves the following steps:
1. Teacher-Student Model Configuration: A DNN serves as the
I. INTRODUCTION teacher model, and an LSM serves as the student model,
The 5th generation (5G) mobile network interconnects minimizing the disparity in model capacity between them.
everyone and everything together, including machines and 2. Loss Functions: Three distinct loss functions train the
student model:
devices. The multi-Gbps data speeds, ultra-low latency, and Output Loss Function: Uses the ground truth label as the
high bandwidth provided by the 5G network enable new target value.
applications such as smart cities, smart factories, and Output Distillation Loss Function: Uses the teacher model’s
autonomous vehicles. A driving force behind the 5G network is predictions as target values.
the utilization of orthogonal frequency division multiplexing Layer Distillation Loss Function: Uses the output of the
(OFDM) in multiple-input multiple-output (MIMO) wireless teacher model’s intermediate layers as target values.
channels. Signal processing in OFDM-MIMO systems is 3. Decoder Implementation: To facilitate communication
between DNNs and SNNs, a decoder translates the SNN's
challenging and resource-consuming, especially in scenarios
intermediate layer spikes into real numbers.
like the massive MIMO architecture for millimeter-wave bands 4. Input Spike Encoding: Converts real-valued inputs into
with high nonlinear distortion in RF components. With the binary spikes using rate coding, which is robust for large-scale
development of deep neural networks (DNNs), NN-based settings.
algorithms for signal processing in OFDM-MIMO systems have 5. Training Process: The student model is trained using a
gained immense attention. These DNN-based algorithms, combined loss function that includes the three loss
however, require a large amount of training data and components. The teacher model is also optimized using output
and output distillation loss functions.
computation/storage resources. Symbol detection in OFDM-
6. Bi-Directional Learning: The DNN adapts to the SNN by
MIMO systems is time-critical and requires low energy learning from the SNN’s output, enabling a mutually beneficial
consumption, making it impractical for DNN-based algorithms to exchange of knowledge.
perform real-time and energy-efficient inference in practical
OFDM-MIMO systems. Therefore, reducing the energy Compared to the bootstrap SNN training algorithm, the DNN-
consumption of MIMO-OFDM systems is beneficial for building SNN co-learning method has three major differences:
Direct Training: Unlike the indirect approach of bootstrap SNN
a sustainable environment. Reservoir computing (RC), derived
training, our method directly trains the SNN model.
from recurrent neural networks (RNNs), is a promising
framework for addressing these challenges. The liquid state Instant Feedback: Our approach provides real-time exchange
machine (LSM), a type of RC that uses spiking neural networks of knowledge and feedback, unlike the delayed feedback in
(SNNs), has two main building blocks: the reservoir layer and bootstrap SNN training.
the readout layer. The reservoir layer maps time-varying input Weight Parameter Handling: Our method involves direct
to high-dimensional spaces and captures both temporal and training of SNN model weights with bi-directional learning,
spatial information, while the readout layer processes this ensuring mutual adaptability and optimization.This DNN-SNN
co-learning algorithm enhances the performance of SNNs by
information. LSMs are easier to train, require less training data, leveraging DNN knowledge, making it suitable for sustainable
and are more energy-efficient compared to DNNs, making them symbol detection in 5G systems on neuromorphic chips.
suitable for symbol detection in OFDM-MIMO systems.
 III. RESULTS The paper demonstrates results on a specific 5G symbol
The experimental setup for symbol detection involved a Liquid detection task, so extending it to more diverse settings
State Machine (LSM) with one reservoir layer of 32 neurons and would be valuable.The encoding and decoding between
two readout layers with 16 and 2 neurons, respectively. The LSM, DNNs and SNNs, using Poisson spike encoding and
which ran on a single Loihi 2 chip using current-based Leaky decoders, could be optimized. Exploring alternative
Integrate-and-Fire (LIF) neurons, was trained with a DNN-SNN co- encoding/decoding schemes may improve the knowledge
learning algorithm. We compared its performance against state-of- transfer between the two model types.
the-art models like LMMSE, DetNet, MMNet, OAMPNet, and
various deep learning models including MLP, CNN, and RNN. For Deploying the co-learned SNN on the Loihi neuromorphic chip
fair comparison, power and energy consumption were for real-time inference would demonstrate the practical benefits of
standardized between the GPU (NVIDIA RTX A2000) and Loihi the approach. Since Loihi is designed for
platforms. Energy measurements indicated that the LSM on Loihi energy-efficient neuromorphic computing, this would
consumed significantly less energy per sample compared to GPU- showcase the system's potential for sustainable
based models, specifically 2.35 times less than QMLP, 5.92 times 5G symbol detection.Comparing the proposed DNN-
less than QRNN, and 3.77 times less than QCNN. In terms of Bit SNN co-learning method to other popular SNN training
Error Rate (BER) for symbol detection in MIMO-OFDM systems, algorithms like surrogate gradient learning or direct
the LSM trained with the DNN-SNN co-learning algorithm (QLSM training of SNNs would provide a more comprehensive
+DSCL) outperformed other models across various Signal-to- evaluation of its performance and tradeoffs.In summary, the
Noise Ratios (SNRs), achieving a 5.7% reduction in BER on paper presents a promising DNN-SNN co-learning framework
average compared to the LSM trained without co-learning (QLSM). for 5G symbol detection on neuromorphic hardware.
These results highlight that our LSM model, when trained with the Future work could explore extensions to the method,
DNN-SNN co-learning algorithm, not only improves accuracy but evaluate it at larger scales, and demonstrate real-
also enhances energy efficiency, making it a promising solution for world deployment on Loihi chips.
sustainable symbol detection in 5G systems. Furthermore, the
ACKNOWLEDGMENT
LSM+DSCL model demonstrated superior BER performance
compared to traditional and deep learning-based symbol detection First and foremost, we express our deepest gratitude to Amma,
algorithms, particularly in environments with varying SNR levels. for her unwavering support, love, and encouragement throughout
this project. We extend our heartfelt thanks to the Head of the
This underscores the potential of using neuromorphic hardware
Department, Dr Ramesh Bhakthavatchalu, for their invaluable
and advanced training algorithms to develop efficient and robust guidance and constantsupport. We are also deeply grateful to the
communication systems. Vice Head of the Department, Remya mam, for their insightful
IV. CONCLUSION AND FUTURE WORK suggestions and encouragement.We sincerely acknowledge
In this study, they have developed a symbol detection algorithm for Prof.Aswathy K Nair, Assistant Professor,Department of
MIMO-OFDM systems using a Liquid State Machine (LSM) Electronics And Communication, AmritaVishwa Vidyapeetham, for
implemented on the Loihi neuromorphic chip. The LSM leverages its her invaluable guidance and supportthroughout this project. We
ability to capture spatial and temporal information from input signals, are deeply indebted to our class advisors Prof.Aswathy K Nair and
enabling it to recover corrupted symbols from distortion and noise at Prof. Baby Sreeja for their constant encouragement and insightful
the receiver. Due to its utilization of a Spiking Neural Network (SNN), feedback. We thank the Department of Electronicsand
the LSM proves to be more energy-efficient compared to traditional Communication and its faculty members for their facilitation and
resources that made this project possible.
DNN-based models.To enhance the accuracy of LSM model, we
introduced a novel DNN-SNN co-learning algorithm. This bi- REFERENCES
directional learning approach facilitates mutual knowledge transfer
[1]H. A. Ammar, R. Adve, S. Shahbazpanahi, G. Boudreau and K. V. Srinivas,
between the DNN and LSM. The DNN provides crucial insights to "User-Centric Cell-Free Massive MIMO Networks: A Survey of Opportunities,
the LSM, improving its ability to mimic complex behaviors modeled Challenges and Solutions," in IEEE Communications Surveys & Tutorials, vol. 24,
by the DNN. Conversely, the LSM enhances the DNN’s adaptability no. 1, pp. 611-652, Firstquarter 2022, doi: 10.1109/COMST.2021.3135119.
by transferring learned insights back to the DNN. This symbiotic keywords: {Massive MIMO;Radio transmitters;Computer
learning process is facilitated through a decoder that converts LSM architecture;Interference;Optical transmitters;Network
architecture;Microprocessors;User-centric cell-free massive MIMO;user-centric
intermediate layer outputs into real numbers for effective cell-free MIMO;cell-free massive MIMO;cooperation;coordination;distributed
communication with the DNN.The experimental results demonstrate massive MIMO;distributed antenna systems;5G/B5G;6G},
the effectiveness of the LSM trained with the DNN-SNN co-learning [2]H. A. Kassir, Z. D. Zaharis, P. I. Lazaridis, N. V. Kantartzis, T. V. Yioultsis and
algorithm. Compared to established symbol detection algorithms T. D. Xenos, "A Review of the State of the Art and Future Challenges of Deep
such as LMMSE, DetNet, MMNet, and OAMPNet, our LSM model Learning-Based Beamforming," in IEEE Access, vol. 10, pp. 80869-80882, 2022,
exhibits a significant reduction in Bit Error Rate (BER). Furthermore, doi: 10.1109/ACCESS.2022.3195299.
in comparison to other DNN-based models like RNN and CNN, our keywords: {Array signal processing;Antenna arrays;Direction-of-arrival
estimation;Interference;Maximum likelihood estimation;Convergence;Support
LSM model achieves approximately six times less energy vector machines;Artificial intelligence;beamforming;deep learning;deep neural
consumption per sample while maintaining comparable networks;direction of arrival estimation;intelligent reflecting surfaces;machine
accuracy.Specifically, LSM model trained with the DNN-SNN co- learning;massive MIMO;MIMO;neural networks},
learning algorithm achieves an average 5.7% improvement in BER [3]J. Wang et al., "3D Beamforming Technologies and Field Trials in 5G Massive
compared to the LSM trained using the surrogate gradient descent MIMO Systems," in IEEE Open Journal of Vehicular Technology, vol. 1, pp.
362-371, 2020, doi: 10.1109/OJVT.2020.3030774.
algorithm. These findings underscore the potential of SNN-based keywords: {Antenna arrays;Array signal processing;Three-dimensional
approaches on neuromorphic hardware like Loihi for achieving displays;Buildings;Wireless communication;Two dimensional
sustainable and efficient symbol detection in advanced displays;5G;Massive MIMO;3D beamforming;multiuser communications},
The paper proposes exploring more advanced loss functions beyond [4]S. Liu, Y. Liang and Y. Yi, "DNN-SNN Co-Learning for Sustainable Symbol
Detection in 5G Systems on Loihi Chip," in IEEE Transactions on Sustainable
the output loss, output distillation loss, and layer distillation loss Computing, vol. 9, no. 2, pp. 170-181, March-April 2024, doi: 10.1109/
functions used to train the student SNN model[1]. Investigating other TSUSC.2023.3324339.
loss functions or combinations of losses could further improve the keywords: {Symbols;Neurons;Artificial neural networks;Adaptation
performance of the co-learning algorithm.Applying the DNN-SNN co- models;OFDM;Task analysis;Computational modeling;Deep learning;deep neural
network;knowledge distillation;machine learning;spiking neural
learning method to larger datasets and deeper network architectures
network;sustainable MIMO symbol detection},
would help validate its scalability and generalization.