Source-linked AI summary
Deep Learning for Wireless Communications: An Emerging Interdisciplinary Paradigm
Linglong Dai, Ruicheng Jiao, Fumiyuki Adachi, H. Vincent Poor, Lajos Hanzo
TL;DR
Wireless communications face unknown channel models, colored noise, and low-latency demands that challenge conventional methods. This review examines DL-based architecture and algorithm design, finding that DL can support complex communication scenarios and accelerate processing while maintaining performance.
Problem
Unknown channel models, colored noise, and latency requirements in emerging wireless systems challenge conventional communication methods.
Method
The review analyzes DL-based architecture design and DL-based algorithm design through wireless communication examples, including unfolded AMP, joint transceiver design, and algorithm mimicry.
Results
DL can support communication systems in complex operating scenarios and speed large-scale processing with guaranteed performance, while BP-CNN reduces iterations needed for the same BER.
Takeaways & Limitations
DL enables system-level replacement of conventional communication blocks and flexible physical-layer designs that can reduce processing latency.
Takeaways & Limitations
DL-based wireless methods lack solid mathematical foundations for theoretical analysis, and their training strategies and network structures remain insufficiently interpretable.
Abstract
from arXiv · showhide
Wireless communications are envisioned to bring about dramatic changes in the future, with a variety of emerging applications, such as virtual reality (VR), Internet of things (IoT), etc., becoming a reality. However, these compelling applications have imposed many new challenges, including unknown channel models, low-latency requirement in large-scale super-dense networks, etc. The amazing success of deep learning (DL) in various fields, particularly in computer science, has recently stimulated increasing interest in applying it to address those challenges. Hence, in this review, a pair of dominant methodologies of using DL for wireless communications are investigated. The first one is DL-based architecture design, which breaks the classical model-based block design rule of wireless communications in the past decades. The second one is DL-based algorithm design, which will be illustrated by several examples in a series of typical techniques conceived for 5G and beyond. Their principles, key features, and performance gains will be discussed. Furthermore, open problems and future research opportunities will also be pointed out, highlighting the interplay between DL and wireless communications. We expect that this review can stimulate more novel ideas and exciting contributions for intelligent wireless communications.
I. INTRODUCTION
Emerging wireless applications expose limits in model-based design, especially when environments are difficult to model or systems require large-scale, low-latency processing. This review organizes DL applications around architecture and algorithm design while highlighting their interaction.
- Motivation: Complex environments and large-scale networks challenge model-based wireless design because channel behavior may be difficult to describe accurately.Examples include underwater environments, hardware nonlinearities, massive MIMO, and massive IoT.
- Motivation: DL can learn intricate variable relationships without accurate mathematical models and can offer computational efficiency through lightweight or parallelized DNNs.These properties motivate DL for large-scale wireless problems.
- Review scope: The review proposes DL-based architecture design and DL-based algorithm design as complementary methodologies for wireless communications.Architecture design includes receiver and joint transceiver design, while algorithm design addresses transmission and optimization processing.
- Review scope: The paper discusses principles, key features, performance gains, open problems, and future opportunities concerning the interplay between DL and wireless communications.The paper is organized around relationship, architecture, algorithm, challenge, and conclusion sections.
II. DIFFERENT PARADIGMS OF WIRELESS COMMUNICATIONS AND DL
Wireless transmission and DL differ in their reliance on mathematical models, modularity, interpretability, and application scope. The paper seeks to combine their complementary benefits while avoiding their weaknesses.
- Paradigm differences: Wireless transmission typically uses accurate mathematical models and separately optimized modules, whereas DL trains DNN parameters jointly and can operate without such models.The contrast includes model dependence, design granularity, and training style.
- Paradigm differences: Classical wireless methods are often intuitive and broadly applicable, while DL methods can be non-intuitive, difficult to interpret, and application-specific.Different tasks may require separately trained neural networks.
- Interplay: The review aims to amalgamate the benefits of DL and wireless transmission while circumventing their weaknesses.It presents the interplay in both directions: DL offers a new wireless design paradigm, while wireless expertise can inform DL techniques.
III. DL-BASED ARCHITECTURE DESIGN FOR WIRELESS COMMUNICATIONS
DL-based architecture design applies neural networks to receiver components and broader communication-system design. The reviewed receiver approach jointly handles difficult signal-processing blocks when nonlinear effects or unknown channel behavior undermine conventional modeling.
- DL-Based Receiver Design: DL-based receiver design jointly optimizes several receiver blocks and can improve BER when nonlinear effects are encountered.This section places receiver design before joint transmitter–receiver design.
- DL-Based Receiver Design: Block-based wireless design separates functions such as source and channel coding, but mathematical-model dependence can limit accuracy in excessively complex scenarios.Channel estimation and equalization may incur detection losses when the equivalent channel is difficult to describe.
- DL-Based Receiver Design: A DL-based OFDM receiver uses a DNN to learn channel behavior and jointly design channel estimation, equalization, and decoding under nonlinear noise.The receiver uses ReLU hidden layers and a sigmoid output layer for signal detection.
B. DL-Based Transceiver Design
DL-based transceiver design replaces separately optimized communication blocks with an end-to-end autoencoder spanning transmitter and receiver. The resulting DNN system is evaluated through BLER comparison with a traditional block-based system.
- B. DL-Based Transceiver Design: Joint transceiver design treats the complete point-to-point system as a single end-to-end autoencoder rather than optimizing receiver components alone.The transmitter encodes messages into normalized signals, a Gaussian-noise layer models the channel, and the receiver reconstructs symbols.
- B. DL-Based Transceiver Design: The DNN-based communication system is contrasted with the classic block-based system through block error rate comparison.Figure 2 specifies BLER as the comparison metric between the two systems.
- B. DL-Based Transceiver Design: The architecture can replace all baseband components with one DNN trained as a whole using cloud computation and storage resources.This whole-system optimization differs from isolated block optimization.
IV. DL-BASED ALGORITHM DESIGN FOR WIRELESS COMMUNICATIONS
DL-based algorithm design uses deep learning to accelerate wireless-communication processing while maintaining reliable performance. The paper illustrates this methodology through transmission algorithms, including sparse channel estimation for mmWave massive MIMO.
- DL-based algorithm design accelerates wireless-communication processing through parallel architectures while potentially improving performance and reducing latency.The methodology covers transmission and optimization algorithms for 5G and beyond, including mmWave massive MIMO, LDPC, NOMA, and UDN.
- Sparse mmWave massive-MIMO channel estimation requires accurate CSI, but iterative AMP incurs repeated linear updates and nonlinear shrinkage operations.The mmWave channel is sparse in the angular domain, motivating compressive-sensing methods such as AMP.
- DL-based AMP unfolds each AMP iteration into one DNN layer, representing residual updates with linear operations and shrinkage with an activation function.Unlike classical AMP, the unfolded method learns shrinkage parameters rather than fixing them in advance.
2) DL-Based Belief Propagation for LDPC Decoding:
The BP-CNN decoder combines standard belief propagation with a CNN to estimate colored noise and iteratively refine LDPC decoding. It achieves better BER than standard BP under the reported simulation setting, implying fewer iterations for the same BER.
- LDPC decoding can suffer from colored noise caused by filtering, oversampling, and multi-user interference in practical systems.LDPC coding was selected for 5G broadband operation because of its strong AWGN performance, but practical detection effects contaminate received signals.
- The decoder subtracts the initial signal estimate from y, refines the resulting noise estimate with a CNN, and feeds the corrected signal back to BP.Iterating between BP and CNN produces a purer signal for subsequent decoding rounds and mitigates colored-noise effects.
- The BP-CNN decoder outperforms standard BP in BER with 50 BP iterations, so matching that BER requires fewer iterations and lower decoding latency.The reported design principle may also apply to turbo and BCH codes because it does not rely on a specific coding scheme or channel model.
B. DL-Based Optimization Algorithm Design
DL-based optimization design targets the high iteration counts and latency of wireless-communication optimization algorithms. The paper presents DNN-based acceleration of power allocation while retaining reliable performance.
- Wireless optimization algorithms can require many iterations to converge, causing high complexity and latency when problem scale is large.This motivates using DL to accelerate processing while maintaining reliable performance.
- WMMSE is widely used for NOMA power allocation, but its iterative convergence makes low-latency operation difficult.The setting considers power allocation for emerging 5G NOMA solutions.
- A DNN mimics WMMSE by learning the mapping from system parameters to power-allocation results, enabling faster testing-time solutions.The approach is designed for interference-contaminated wireless scenarios with K single-antenna transceiver pairs and per-pair power constraints.
2) DL-Based Energy Consumption Minimization for UDN:
For UDN energy minimization, DL reduces the scale of a large optimization problem by predicting variables that can be eliminated. The resulting reduced problem preserves nearly the original performance while shortening solution time.
- UDNs use many access points to provide high data rates, making energy minimization under flow-demand constraints an important large-scale optimization problem.Each access point can establish a transmission link to another access point within its range.
- A DL-based algorithm directly reduces optimization scale by predicting and excluding zero variables from the problem.This addresses a gap left by methods focused mainly on reducing computational complexity rather than problem size.
- The elimination-based method is not tied to a specific scenario or algorithm and can extend to other optimization problems by excluding zero variables.
V. CHALLENGES AND RESEARCH OPPORTUNITIES
DL-based wireless methods remain at an early stage, with theoretical analysis still lacking the mathematical foundations needed to guide design and characterize performance limits.
- DL-based wireless methods are still in their infancy, leaving many challenging issues for further study.
- Theoretical analysis lacks solid mathematical foundations for guiding practical design and revealing performance limits.
2) Interpretability:
Interpretability remains limited because training strategies and network structures lack clear mathematical or physical explanations, while wireless data are difficult to collect, process, and generalize across changing environments.
- Interpretability: Training strategies remain somewhat haphazard, and the reasons certain network structures perform well are difficult to explain physically.
- Interpretability: Advanced optimization theory may help identify more appropriate loss functions and training strategies.
- Interpretability: High-quality wireless training data are difficult to obtain because generating large labeled datasets is challenging and communication data are high-dimensional.
- Interpretability: Wireless datasets are heterogeneous and require mature processing flows, while differing training and testing datasets can produce different performance.
- Interpretability: Rapidly changing channels may require complete retraining from scratch, motivating learning algorithms with greater environmental adaptability and robustness.
D. Implementation-Oriented Challenges
Deploying DL-based methods in real communication systems requires additional infrastructure alongside existing RF and signal-processing capabilities, while the review identifies broader gains and research opportunities.
- D. Implementation-Oriented Challenges: Existing communication infrastructure is mainly equipped for RF and signal processing, so DL deployment may require cloud servers.
- D. Implementation-Oriented Challenges: The review concludes that DL can support communication systems in complex scenarios and speed up large-scale processing with guaranteed performance.
- D. Implementation-Oriented Challenges: Further study may promote DL development in theory and algorithmic innovation.