PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep Learning in Physical Layer Communications

Zhijin Qin, Hao Ye, Geoffrey Ye Li, Biing-Hwang Fred Juang

arXiv:1807.11713v3cs.IT

TL;DR

Physical-layer communications face practical imperfections and locally optimized blocks that may not achieve global optimality. This article surveys data-driven and model-driven DL approaches across block-structured and end-to-end systems, highlighting their reported implementation benefits and future research directions.

  • Problem

    Physical-layer communications must address analytically difficult practical imperfections and communication blocks whose local optimization does not guarantee global optimality.

  • Method

    The article overviews DL applications in physical-layer communications, covering model-driven and data-driven systems with block structures and end-to-end architectures.

  • Results

    The article identifies DL applications in signal compression and signal detection and discusses emerging end-to-end communication systems.

  • Takeaways & Limitations

    DL-based physical-layer communications are presented as a route toward intelligent systems adapted to practical communication settings and future research directions.

Abstract

from arXiv · show

Deep learning (DL) has shown the great potentials to break the bottleneck of communication systems. This article provides an overview on the recent advancements in DL-based physical layer communications. DL can improve the performance of each individual block in communication systems or optimize the whole transmitter/receiver. Therefore, we categorize the applications of DL in physical layer communications into systems with and without block structures. For DL-based communication systems with block structures, we demonstrate the power of DL in signal compression and signal detection. We also discuss the recent endeavors in developing end-to-end communication systems. Finally, the potential research directions are identified to boost the intelligent physical layer communications with DL.

I. INTRODUCTION

The introduction frames physical-layer communications as a gap between analytical models and practical imperfections, and between locally optimized blocks and global optimality. It presents DL as a data-driven or model-driven approach spanning block-structured and end-to-end systems.

  • Challenges: Physical-layer systems face unknown real-world effects that are difficult to express analytically, motivating more adaptive frameworks.Examples include underwater acoustic channels and molecular communications.
  • Challenges: Traditional communication systems optimize processing blocks locally, so global optimality is not guaranteed and structures may vary across environments.The cited blocks include channel encoding, modulation, and signal detection.
  • DL approach: DL can optimize networks or systems over large datasets without requiring a mathematically tractable model.This data-driven property is presented as a response to practical modeling challenges.
  • DL approach: Communication models can be combined with DL to form model-driven systems that use both model-based algorithms and DL.This approach is positioned alongside pure data-driven optimization.
  • Reported benefits: Learned algorithms can execute faster with lower power consumption than manually programmed counterparts because neural networks support parallelization and low-precision data types.The passage attributes these implementation advantages to concurrent architectures and low-precision computation.
  • Scope: The article surveys DL gains for block-structured wireless systems and end-to-end systems that merge communication blocks, then identifies future research directions.The organization covers block-based applications, end-to-end communications, and directions for intelligent physical-layer communications.

II. DEEP NEURAL NETWORKS AND DEEP LEARNING BASED COMMUNICATIONS

This section introduces neural-network foundations and describes how deep-learning architectures support communication applications. It covers feedforward, recurrent, and convolutional networks, along with training and relevant limitations.

  • Deep Neural Networks Basis: A neuron applies a nonlinear activation function to the sum of weighted inputs and bias, and connected neurons form layered architectures.The passage represents the output as y = σ(sum of weighted inputs + b).
  • Deep Neural Networks Basis: A neural network adjusts weights using labelled input-output pairs by minimizing a loss function.Common losses include mean-squared error and categorical cross-entropy, and stochastic gradient descent is used to optimize weights.
  • Deep Neural Networks Basis: Deep neural networks use multiple fully connected hidden layers, with each layer representing a different feature of the input data.Feedforward and recurrent neural networks are presented as typical DNN models.
  • Deep Neural Networks Basis: Deep convolutional networks reduce trainable parameters by retaining only some connections between neurons and adjacent layers.The section also identifies their potential for signal compression and recovery.
  • Deep Neural Networks Basis: Recurrent neural networks use current inputs and previous hidden states, giving them memory for sequential processing.Long short-term memory networks are described as a specialized RNN used in applications including joint source-channel coding.
  • Deep Neural Networks Basis: Long-term temporal dependence in RNNs can produce non-stationary errors during training.LSTM is introduced as a special RNN intended to eliminate some unnecessary information in the network.

2) Generative Adversarial Net (GAN) and Conditional GAN:

GANs generate data resembling a target distribution through an adversarial generator–discriminator game, while conditional GANs add conditioning information. In communications, these models can learn channel-output distributions and provide surrogate channels for transmitter training.

  • GANs can produce data following a target distribution, lowering the amount of labelled data required.
  • GANs use a generator, G, and discriminator, D, in a min-max game over real and generated data.D distinguishes real from fake data, while G produces plausible data to fool D.
  • At the game's outcome, G generates data with the same distribution as real data, preventing D from distinguishing them.
  • A conditional GAN extends GAN by supplying conditioning information, m, to both G and D as an additional input.
  • In communication systems, GANs and conditional GANs model channel-output distributions and can act as surrogate channels during transmitter training.The learned surrogate allows gradients to pass through to the transmitter.

B. Bayesian Optimal Estimator

Bayesian optimal estimation frames communication signal-processing modules as posterior inference problems, but exact estimators are computationally intractable. Model-driven deep learning incorporates physical knowledge and unfolds iterative methods to reduce learned parameters and improve training efficiency.

  • B. Bayesian Optimal Estimator: MIMO detection and channel decoding can be formulated as posterior probability inference in probabilistic graphical models.Received signals are observations, while transmitted signals are latent variables whose dependence is represented explicitly.
  • B. Bayesian Optimal Estimator: Bayesian optimal estimators are obtained by minimizing expected MSE using the posterior distributions of latent variables.
  • B. Bayesian Optimal Estimator: Exact computation of Bayesian optimal estimators is computationally intractable.
  • B. Bayesian Optimal Estimator: Approximate message passing and expectation-propagation efficiently approximate Bayesian-optimal performance, and unfolding can further improve these detectors.Unfolded iterative procedures use DL models whose parameters are updated from training data.
  • C. Deep Learning based Communications: Data-driven DL can replace individual processing blocks or merge them into end-to-end transmitter or receiver networks.These systems use DNN parameters tuned with large amounts of labelled data.
  • C. Deep Learning based Communications: Fully connected DNN training may require substantial time and large datasets, while wireless systems often have scarce computing resources and labelled data.
  • C. Deep Learning based Communications: Model-driven DL exploits channel models, information theory, and other domain knowledge to reduce learned parameters and improve training efficiency.This approach is contrasted with data-driven models that use general structures, many unknown parameters, and large datasets.
  • C. Deep Learning based Communications: Prior knowledge can reduce sample complexity, but inaccurate priors may hurt real-scenario performance; data-driven models require more samples but may be more robust across circumstances.

III. DEEP LEARNING BASED BLOCK-STRUCTURED COMMUNICATIONS

The paper surveys DL applications within block-structured communication systems, emphasizing intelligent signal compression and detection. DL-based methods address feedback overhead, recovery speed, imperfect sparsity, and difficult channel conditions.

  • DL applications in block-structured systems are categorized into intelligent signal compression and signal detection.
  • Intelligent Signal Compression: Large antenna arrays create excessive CSI feedback overhead, motivating compression methods that exploit spatial, temporal, and sparse CSI structure.
  • Intelligent Signal Compression: Traditional CS is limited because real-world data is not exactly sparse and existing recovery algorithms converge relatively slowly.
  • Intelligent Signal Compression: DL-based recovery methods include DCN and LDAMP, with LDAMP offering strong accuracy and speed while CS-based CSI estimation still faces performance limits.
  • Model-Driven CSI Feedback Compression and Reconstruction: CsiNet uses an encoder-decoder architecture to learn CSI compression, feedback, and reconstruction from angular-delay-domain channel matrices.
  • Model-Driven CSI Feedback Compression and Reconstruction: CsiNet remarkably outperforms traditional CS-based methods in both compression ratio and reconstruction performance.

2) Data-Driven Joint Source-Channel Coding:

Data-driven joint source-channel coding uses neural encoders and decoders to preserve semantic information while reducing reliance on separately optimized traditional coding blocks. The surveyed approaches report lower word-error rates and improved coding performance in supported settings.

  • Source coding removes redundancy, whereas channel coding adds redundancy to improve robustness to noise; Shannon separation permits separate design without loss of optimality.
  • Separate source and channel coding can be impractical when communication blocks cannot be made very large, motivating joint source-channel coding.
  • Data-Driven Joint Source-Channel Coding: A DL-based joint source-channel encoder and decoder can output different sentences while preserving their semantic information content.
  • Data-Driven Joint Source-Channel Coding: The proposed text-processing model uses an RNN encoder, binarization layer, channel layer, and RNN decoder with stacked bidirectional LSTMs.
  • Data-Driven Joint Source-Channel Coding: Word-error rate is lowered relative to traditional separate source-channel coding baselines using Huffman and Reed-Solomon coding.
  • DL can also apply source and channel coding separately and achieve better performance than typical coding techniques.

B. Intelligent Signal Detection

DL-based detection can jointly optimize receiver tasks and learn from difficult channel conditions that analytical models characterize poorly. A DNN-based OFDM receiver implicitly infers channel information while directly reconstructing transmitted data.

  • DL-based detection is especially useful when traditional receiver blocks require joint optimization or channels are difficult to characterize analytically.
  • Data-Driven Joint Channel Estimation and Signal Detection: Traditional receivers estimate CSI from pilots before detecting transmitted symbols, whereas joint DL methods combine channel estimation and signal detection.
  • Data-Driven Joint Channel Estimation and Signal Detection: A five-layer fully connected DNN embedded in an OFDM receiver treats the channel as a black box and takes received pilot and data signals as inputs.
  • Data-Driven Joint Channel Estimation and Signal Detection: The DNN implicitly infers channel information and predicts transmitted data directly without explicitly estimating the channel.
  • Data-Driven Joint Channel Estimation and Signal Detection: The DNN-based method outperforms the minimum MSE-based approach without adequate pilots or cyclic prefix and under nonlinear distortion.
  • Supervised parameter updates allow the data-driven method to address adversarial effects and improve robustness to undesired circumstances.

2) Model-Driven MIMO Detection:

Model-driven MIMO detection incorporates learning into iterative algorithms to improve adaptability while retaining a compact trainable structure. The surveyed approach improves orthogonal AMP detection performance with limited training demands and potential massive-MIMO scalability.

  • Analytical detectors often assume channel distributions, limiting performance in complicated environments; learning-based parameters can adapt to specific data.
  • Model-Driven MIMO Detection: Model-driven detection unfolds iterative procedures into a signal-flow graph and tunes only several critical variables through supervised learning.
  • Model-Driven MIMO Detection: The trainable orthogonal AMP framework sets only two variables as trainable parameters in each iteration.
  • Model-Driven MIMO Detection: The model-driven detector trains in a shorter period and with less training data than a regular DNN while improving orthogonal AMP detector performance in Rayleigh and correlated MIMO channels.
  • Model-Driven MIMO Detection: The approach has potential for scaling to massive MIMO communications.
  • DL has been applied across physical-layer tasks including channel decoding, beam prediction, and molecular signal detection, with reported robustness or contextual advantages.

IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS

End-to-end learning treats the communication system as a single reconstruction task, allowing transmitter and receiver designs to be optimized jointly without conventional block structures. The section addresses unknown-channel gradients through reinforcement learning and related methods.

  • IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS: The section introduces two methods to address the missing-gradient issue in learning-based end-to-end communication systems.These methods are presented after identifying the unknown-channel obstacle to transmitter optimization.
  • IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS: End-to-end learning jointly trains transmitter and receiver to encode and recover source data using a supervised loss on recovery accuracy.This removes the requirement for conventional communication-system block structures.
  • IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS: Unknown channel parameters block gradient back-propagation to the transmitter, while assumed channel models can bias the learned system.Real channels may include unknown or analytically inexpressible noise and time-varying effects.
  • IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS: Reinforcement learning treats the transmitter as an agent and the channel plus receiver as the environment, using end-to-end loss as feedback.The transmitter maps source-data states to transmit-signal actions and optimizes cumulative rewards with policy gradients.
  • IV. DEEP LEARNING BASED END-TO-END COMMUNICATIONS: Policy-gradient reinforcement learning optimizes the end-to-end loss without requiring gradients from the channel.The receiver calculates the sample loss and returns it as an environmental reward guiding transmitter training.

B. Conditional GAN Based End-to-End Systems

Conditional GANs enable end-to-end communication learning without prior channel information by modeling channel effects and providing a surrogate through which transmitter gradients can pass. Reported simulations show performance similar to Hamming (7,4) coding with maximum-likelihood decoding.

  • B. Conditional GAN Based End-to-End Systems: A conditional GAN enables end-to-end communication learning without prior channel information by modeling the conditional distribution of the channel.The pipeline includes transmitter, channel-generator, and receiver DNNs trained iteratively.
  • B. Conditional GAN Based End-to-End Systems: The conditional GAN mimics channel effects as a surrogate channel, enabling gradients to pass through and train the transmitter.Its conditioning uses transmit signals together with received pilot information, making the generated distribution specific to the instantaneous channel.
  • B. Conditional GAN Based End-to-End Systems: Similar performance as the Hamming (7,4) code with maximum-likelihood decoding (MLD) was reported for the conditional GAN end-to-end system.The simulation results were presented as evidence of the system’s effectiveness.
  • B. Conditional GAN Based End-to-End Systems: The article summarizes DL applications in physical-layer communications and identifies research directions for bringing intelligent communications from theory to practice.The discussion places end-to-end systems within the paper’s broader overview and future-work agenda.
  • B. Conditional GAN Based End-to-End Systems: Initial results indicate that DL-based end-to-end communications are comparable with traditional communication systems.The paper identifies unresolved questions about eventual performance, complexity, and achievable gain.

B. Tradeoff between System Performance and Training Efficiency

End-to-end designs can optimize a universal transmitter and receiver, but merging all communication blocks makes training very long. The paper therefore discusses balancing training efficiency with system performance and adapting metrics to application-specific goals.

  • B. Tradeoff between System Performance and Training Efficiency: End-to-end learning can optimize a universal transmitter and receiver, but merging all communication blocks makes the training process very long.Keeping some blocks and using model-driven DL methods is proposed as a way to improve training efficiency and system performance.
  • B. Tradeoff between System Performance and Training Efficiency: System design must trade off training efficiency against communication-system performance when retaining blocks or applying model-driven DL.The paper calls for careful system design to achieve this balance.
  • C. Communication Metric Learning: Real applications may require information sharing objectives that differ from error-free reconstruction of transmitted data.The paper notes that reconstruction metrics may not be satisfactory for every task.
  • C. Communication Metric Learning: Bit-error rate is not a good metric for image and video transmission because it cannot reflect human visual perception.End-to-end systems should revise their metrics to address application-specific requirements.
  • C. Communication Metric Learning: Application-oriented communication may tolerate transmission errors when the semantic information needed by the receiver’s task remains intact.The approach treats transmitted data as unequally important and prioritizes preserved semantic information.
  • B. Tradeoff between System Performance and Training Efficiency: Wireless-communication datasets remain underdeveloped, while data-protection and privacy regulations limit open access to real-world communication data.The paper identifies publishing channel responses under different conditions as important for accelerating DL-based physical-layer research.
Loading 1807.11713v3…