ACCEPTED FROM OPEN CALL

Deep Learning in
Physical Layer Communications
Zhijin Qin, Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang

Abstract ular communications. Therefore, a more adaptive

framework is required to handle the challenges.
DL has shown great potential to revolution- Block Structures versus Global Optimality:
ize communication systems. This article pro- Traditional communication systems consist of sev-
vides an overview of the recent advancements eral processing blocks, such as channel encod-
in DL-based physical layer communications. DL ing, modulation, and signal detection, which are
can improve the performance of each individual designed and optimized within each block local-
block in communication systems or optimize the ly. Thus global optimality cannot be guaranteed.
whole transmitter/receiver. Therefore, we cat- Moreover, the optimal communication system
egorize the applications of DL in physical layer structure varies with environments. As a result, opti-
communications into systems with and without mal or robust communication systems for different
block structures. For DL-based communication scenarios are more than desired.
systems with the block structure, we demonstrate DL could be a pure data-driven method, where
the power of DL in signal compression and signal the networks/systems are optimized over a large
detection. We also discuss the recent endeavors training data set and a mathematically tractable
in developing DL-based end-to-end communica- model is unnecessary. Such a feature motivates us
tion systems. Finally, potential research directions to exploit DL in communication systems in order to
are identified to boost intelligent physical layer address the aforementioned challenges. In this sit-
communications. uation, communication systems can be optimized
for specific hardware configuration and channel
Introduction to address the imperfection issues. On the other
The idea of using neural networks (NN) to intelli- hand, many models in physical layer communica-
gentize machines can be traced to 1942 when a tions have been established by researchers and
simple model was proposed to simulate the status engineers during the past several decades. Those
of a single neuron. Deep learning (DL) adopts a models can be combined with DL to design mod-
deep neural network (DNN) to find data repre- el-driven DL-based communication systems, which
sentation at each layer, which could be built by can take advantage of both model-based algo-
using different types of machine learning (ML) rithms and DL [2].
techniques, including supervised ML, unsuper- There is evidence that the “learned” algorithms
vised ML, and reinforcement learning. In recent could be executed faster with lower power con-
years, DL has shown its overwhelming privilege sumption than the existing manually “programmed”
in many areas, such as computer vision, robot- counterparts, as NNs can be highly parallelized
ics, and natural language processing, due to its on the concurrent architectures and implemented
advanced algorithms and tools in learning compli- with low-precision data types. Moreover, the pas-
cated models. sion on developing artificial intelligence-powered
Different from the aforementioned DL appli- devices from manufacturers, such as Intel© Movid-
cations, where it is normally difficult to find a ius™ Neural Compute Stick, has also boosted the
concrete mathematical model for feature represen- boom of DL-based wireless communications.
tation, various theories and models, from informa- This article will identify the gains that DL can
tion theory to channel modelling, have been well bring to wireless physical layer communications,
developed to describe communication systems [1]. including systems with the block structure and the
However, the gap between theory and practice end-to-end structure merging those blocks. The rest
motivates us to work on intelligent communica- of this article is organized as follows. The follow-
tions. In particular, the following challenges have ing section introduces the important basis of DNN
been identified in existing physical layer commu- and illustrates DL-based communication systems.
nications. We then discuss how to apply DL to block-struc-
Mathematical Model versus Practical Imper- tured communication systems. Following that we
fection: Conventional communication systems rely demonstrate DL-based end-to-end communication
on the mathematically expressed models for each systems, where an individual block for a specific
block, while in real-world applications, complex sys- function, such as channel estimation or decoding,
tems may contain unknown effects that are difficult disappears. The final section concludes this article
to be expressed analytically. For example, it is hard with potential research directions in the area of
to model underwater acoustic channels or molec- DL-based physical layer communications.

Zhijin Qin is with Queen Mary University of London; Digital Object Identifier:
Hao Ye, Geoffrey Ye Li, and Biing-Hwang Fred Juang are with Georgia Institute of Technology. 10.1109/MWC.2019.1800601

IEEE Wireless Communications • April 2019 1536-1284/19/$25.00 © 2019 IEEE 93

 x1
w1
n
x2 y i 1
wi xi b
w2

wn

xn

FIGURE 1. Development of neural networks: a) single neuron element; b) feedforward neural network (FNN); c) recurrent neural net-
work (RNN).

Deep Neural Networks and boosted many applications due to the powerful
algorithms and tools. DCN has shown its great
Deep Learning Based Communications potential for signal compression and recovery
In this section, we will first introduce the basis of problems, which will be demonstrated below.
DNN, generative adversarial network (GAN), con- For the RNN in Fig. 1c, the outputs of each
ditional GAN, and Bayesian optimal estimator, layer are determined by both the current inputs
which are widely used in DL-based communica- and their hidden states in the previous time step.
tion systems. Then we will discuss the intelligent The critical difference between FNN and RNN is
communication systems with DL. that the latter has memory and can capture the
hidden layer outputs in the previous step. How-
Deep Neural Networks ever, as RNN is dependent on time over a long
Deep Neural Networks Basis: As aforementioned, term, non-stationary errors may show up during
research on NN started from the single neuron. As the training process. A special type of RNN, named
shown in Fig. 1a, the inputs of the NN are {x1, x2, long short-term memory (LSTM), has been further
… , xn} with the corresponding weights, {w1, w2, … proposed to eliminate some unnecessary infor-
, wn}. The neuron can be represented by a non-lin- mation in the network. LSTM has been widely
ear activation function, s(•), that takes the sum of applied in various cases, such as the joint deign
the weighted inputs. The output of the neuron can of source-channel coding, which will be briefly dis-
n
be expressed as y = s(Si=1wixi + b), where b is the cussed later.
shift of the neuron. An NN can be established by Generative Adversarial Net (GAN) and Con-
connecting multiple neuron elements to generate ditional GAN: Training a typical DNN is heavily
multiple outputs to construct a layered architec- dependent on the large amount of labelled data,
ture. In the training process, the labelled data, that which may be difficult to obtain or even unavail-
is, a set of input and output vector pairs, is used able in certain circumstance. As shown in Fig. 2,
to adjust the weight set, W, by minimizing a loss GAN is a type of generative method, which can
function. In the NN with single neuron element, produce data that follows a certain target distri-
W = {b, w1, w2, … , wn}. The commonly-used loss bution. By doing so, demands for the amount of
functions include mean-squared error (MSE) and labelled data can be lowered. In Fig. 2, a GAN
categorical cross-entropy. To train the model for a consists of a generator, G, and a discriminator, D.
specific scenario, the loss function can be revised D attempts to differentiate between the real data
by introducing the l1-norm or l2-norm of W or acti- and the fake data generated by G while G tries
vations. The l1-norm or l2-norm of W can also be to generate plausible data to fool D into making
introduced in the loss function as the regularizer to mistakes, which introduces a min-max two player
improve the generalization capabilities. Stochastic game between G and D. As a result of the min-max
gradient descent (SGD) is one of the most popular two player game, the generator, G, will generate
algorithms to optimize W. data with the same distribution as the real data,
With the layered architecture, a DNN includes and therefore the discriminator, D, cannot identi-
multiple fully connected hidden layers, in which fy the difference between the real and fake data.
each of them represents a different feature of the Conditional GAN is an extension of GAN by pro-
input data. Figures 1b and 1c show two typical viding extra conditioning information, m, where the
DNN models: feedforward neural network (FNN) conditioning information has been fed to both G
and recurrent neural network (RNN). In FNNs, and D as the additional input.
each neuron is connected to the adjacent layers In communication systems, a GAN and a con-
while the neurons in the same layers are not con- ditional GAN can be applied to model the distribu-
nected to each other. The deep convolutional net- tion of the channel output. Moreover, the learned
work (DCN) is developed from the fully connected model can be utilized as a surrogate of the real
FNN by only keeping some of the connections channel when training the transmitter so that the
between neurons and their adjacent layers. As a gradients can pass through to the transmitter. An
result, DCN can significantly reduce the number application example of conditional GAN will be
of parameters to be trained [3]. Recently, DL has introduced later.

94 IEEE Wireless Communications • April 2019

 Bayesian Optimal Estimator Real data
Real
Besides the standard DL models, the expert Dataset x
knowledge can be beneficial in modifying the Real
structures of DL models to provide more explain- m data
able and predictable models in physical layer D
communications. In fact, many signal process-
ing modules in communication systems, such as
multi-input multi-output (MIMO) detection and
channel decoding, can be cast as posterior proba- Input noise Fake
G Fake data
bility inference problems in probabilistic graphical z data
models, where the dependence of observation
variables (e.g., the received signals) and the latent Conditioning
variables (e.g., the transmitted signals) are information m
expressed explicitly. The posterior distribution of m
the latent variables can be calculated via Bayes-
ian inference and then the Bayesian optimal esti- FIGURE 2. Structure of conditional GAN [4].
mators are obtained by minimizing the expected
MSE with given posterior distributions.
Even though the exact computation of the Bayes-
ian optimal estimators are computationally intracta-
ble in many problems, some iterative approaches,
such as approximate message passing (AMP) and
expectation-propagation, can approximate the per-
formance of Bayesian optimal efficiently. Moreover,
these iterative detectors can be further improved
by unfolding and representing the iterative proce-
dures with DL models, where the parameters of the
model can be updated based on the training data. A
detailed example will be shown later.
FIGURE 3. Intelligent communication system structure.
Deep Learning-based Communications
Figure 3 shows the intelligent communication
system structure with DL. Compared to the con- a huge data set. Each of them has its advantages
ventional block-based communication structure, and disadvantages. In general, it involves the trade-
DL-based communication systems utilize the off between variance and bias in the learning the-
power of DL to facilitate transmission. A data-driv- ory. With prior knowledge, the sample complexity
en DL-based communication system is usually for learning models can be largely reduced, but the
represented by a DNN, and a large amount of models may suffer when the prior knowledge is not
labelled data is used to tune the parameters of accurate in the real scenario. On the contrary, the
the DNN. Such a DNN can be regarded as a data-driven model is with less presumption. The sam-
“black box” and used in each processing block ple complexity is large but it can be more robust
individually to replace the existing algorithms, under variant circumstances.
which is the overlap of block-structured and
data-driven modules shown in Fig. 3. Moreover,
it can be used to optimize the whole transmitter
Deep Learning-based
or the whole receiver when a DNN merges all Block-Structured Communications
processing blocks in the communication system. Even though the existing block-structured com-
Such a structure is called an end-to-end communi- munication systems have been carefully designed
cation system. Examples of data-driven DL-based from their infancy to the fifth generation (5G),
communication systems will be introduced below. more efforts are still required to break the bottle-
Typically, training a fully connected DNN neck in wireless communication systems. In this
requires plenty of training time in addition to a section, we focus on the applications of DL in
huge data set, especially in end-to-end communi- different communication blocks, which are cat-
cation systems. However, the computing resourc- egorized into intelligent signal compression and
es and labelled data are often scarce in wireless detection.
communication systems. As shown in Fig. 3, mod-
el-driven DL methods exploit the known physi- Intelligent Signal Compression
cal mechanism and domain knowledge, such as Most types of source data exhibit unique inner
well-developed channel models and information structures that can be utilized for compression.
theory, which can reduce the number of param- Such structured data can be modelled by differ-
eters to be learned and improve the training effi- ent approaches. Sparse representation is a com-
ciency of some block-structured communication monly-used one. It is worth noting that the most
systems. An example of model-driven DL based important property of DL is that it can automat-
wireless communications will be provided below. ically find compact low-dimensional representa-
The model-driven methods exploit some prior tions/features of high dimensional data [5, 6],
knowledge of the system to reduce the number which can be demonstrated by the following two
of parameters to be learned, while the data-driven examples.
methods assume a general system structure that usu- Model-Driven CSI Feedback Compression
ally has lots of unknown parameters to be trained by and Reconstruction: In the downlink of frequency

IEEE Wireless Communications • April 2019 95

Most types of source
data exhibit unique CSI encoder at user side CSI decoder at base station

inner structures that Nx1 Nx1 RefineNet

Compressed CSI
can be utilized for Mx1 Mx1
compression. Such

Noise layer

RefineNet
structured data can be Channel Recovered
matrix CSI
modelled by different
approaches. Sparse
representation is a Convolutional Dense Dense Convolutional
commonly-used one. It layer layer
is worth noting that the
most important prop- FIGURE 4. DL-based channel compression, feedback, and recovery by CsiNet [10].
erty of DL is that it
can automatically find
division duplex networks, massive MIMO relies while channel coding improves the robustness to
compact low-dimen- on channel state information (CSI) feedback to noise by adding redundancy to the coded informa-
sional representations/ achieve performance gains from multiple antennas tion when it is transmitted over channels. The Shan-
features of high dimen- at the base station. However, the large number of non separation theorem guarantees that source
sional data. antennas results in excessive feedback overhead. coding and channel coding can be designed sepa-
Extensive work has been carried out to reduce the rately without loss of optimality. However, in many
feedback overhead by utilizing the spatial and tem- communication systems, source coding and chan-
poral correlations of CSI. By exploiting the sparse nel coding are designed jointly as it is not practical
property of CSI, compressive sensing (CS) has to have very large blocks.
been applied to compress CSI at the user side and A joint source-channel coding based on DL has
the compressed CSI is then recovered at the base been proposed in [11]. With text as the source data,
station. However, traditional CS algorithms face the DL-based source-channel encoder and decoder
challenges as real-world data is not exactly sparse may output different sentences but preserve their
and the convergence speed of the existing signal semantic information content. Specifically, the pro-
recovery algorithms is relatively slow, which has posed model adopts an RNN encoder, a binariza-
limited the practical applications of CS [7]. tion layer, a channel layer, and an RNN decoder.
DCN has been applied to learn the inverse The text is structured before it is processed by the
transformation from measurement vectors to sig- stacked bidirectional LSTM networks. Then the bina-
nals to improve the recovery speed in CS [7]. In rizer is adopted to output binary values, which are
particular, DCN has two distinctive features that taken as the inputs of the channel. At the receiver, a
make it uniquely applicable to sparse recovery stack of LSTM is used for decoding. By doing so, the
problems. One is that the neurons are sparsely word-error rate is lowered compared with various
connected. The other is with shared weights across traditional separate source-channel coding baselines,
the entire receptive fields of one layer, DCN can such as using Huffman and Reed-Solomon code
increase the learning speed compared to a ful- for source and channel coding, respectively. Even
ly-connected network [8]. Learned denoising-based though this design is particularly for text processing,
AMP (LDAMP) [9] is one of the excellent signal it inspires us to apply DL to where recovery of the
recovery algorithms in terms of both accuracy and exact transmitted data is not compulsory as long
speed, which has been applied to channel estima- as the main information within it is conveyed. For
tion in millimeter-wave (mmWave) communica- example, in sparse support detection, we need to
tions [2]. However, the achieved improvement still determine if there is a sparse support at each loca-
cannot boost the CS-based CSI estimation. tion while the exact amplitude of each location is
CsiNet [10] has been proposed to mimic the not of interest.
CS processes for channel compression, feedback, In addition to the aforementioned two exam-
and reconstruction. In particular, an encoder is used ples, DL has also been widely applied in other sig-
to collect compressed measurements by directly nal compression applications. For example, instead
learning channel structures from the training data. of performing joint source-channel coding, DL can
As shown in Fig. 4, by taking the channel matrix in be applied to source coding and channel coding,
the angular-delay domain as inputs, the first layer of separately, to achieve better performance com-
the encoder is a convolutional layer that generates pared to typical coding techniques. Moreover,
two feature maps. Then the feature maps are vector- DNN has also been widely applied to facilitate the
ized and a fully connected layer is used to generate design of measurement matrix and signal recovery
the real-valued compressed CSI. Only those com- algorithm in CS [6], which can be used in various
pressed CSI is fed back to the base station. With wireless scenarios, that is, channel estimation and
such an encoder, the feedback overhead is signifi- wideband spectrum sensing.
cantly reduced. At the base station, the decoder
reconstructs the CSI via learning the inverse trans- Intelligent Signal Detection
formation of the compressed feedback. It has been DL-based detection algorithms can significantly
shown that CsiNet remarkably outperforms the tra- improve the performance of communication sys-
ditional CS-based methods in terms of both com- tems, especially when the joint optimization of
pression ratio and recovery speed. the traditional processing blocks is required and
Data-Driven Joint Source-Channel Coding: when the channels are hard to be characterized
Typical source coding mainly utilizes the sparse by analytical models. Here, we provide two exam-
property to remove the redundancy in source data ples for DL-based detection.

96 IEEE Wireless Communications • April 2019

 Data-Driven Joint Channel Estimation and
Signal Detection: Traditionally, channel estima-
tion and signal detection are two separate pro-
cedures at the receiver. The CSI is first estimated
by means of pilots prior to the detection of the
transmit symbols. Then with the estimated CSI, the
transmit symbols can be recovered at the receiv-
er. A joint channel estimation and signal detection
approach has been proposed in [12]. Specifical-
ly, a five-layer fully connected DNN is embedded
into an orthogonal frequency-division multiplexing
(OFDM) receiver for joint channel estimation and
detection by treating the channel as a “black box.”
The DNN is trained to reconstruct the transmit data
by feeding the received signals corresponding to
the transmit data and pilots as inputs. Therefore,
the channel information can be inferred implicit-
ly by the DNN and used to predict the transmit
data directly without explicitly estimating the CSI.
Figure 5 demonstrates that the DNN-based chan-
nel estimation and detection method outperforms
the minimum MSE-based approach when without
adequate pilots or cyclic prefix, and with nonlin- FIGURE 5. Performance comparison of DL and minimum MSE-based joint
ear distortion.1 The advantage is that when these channel estimation and signal detection in OFDM systems [11].
adversarial effects occur, the data-driven method
can learn to deal with these effects in a supervised
manner, that is, updating the parameters to mini- taken into consideration to improve the prediction.
mize the empirical loss, which improves the robust- Moreover, DL has shown its privilege on molecular
ness to the undesired circumstances. signal detection when the channel models are opti-
Model-Driven MIMO Detection: In MIMO mized based on training data instead of any prior
detection, iterative methods, which are based on channel information.
Bayesian optimal detectors, have shown superior
performance with moderate computation com-
plexity. However, these detectors often impose
Deep Learning-based
assumptions on the channel distribution, which End-to-End Communications
limits the performance under many complicated In the previous section, we discussed the appli-
environments. By incorporating learning based cations of DL in each individual block of commu-
approaches, the adaptability of the detectors can nication systems. In this section, we will present
be improved since the parameters of the model innovative learning-based communication systems
can be refined according to the specific data. In by treating the entire communication system as
[2], the iterative procedures are unfolded to a sig- an end-to-end reconstruction task [4, 14, 15]. In
nal flow graph. Only several critical variables are particular, based on the data-driven methods, the
required to tune the graph in the supervised learn- transmitter learns to encode the source data into
ing manner. This trainable framework has been encoded symbols (or transmit signals) to be trans-
combined with the orthogonal AMP detector, mitted over the channel while the receiver learns
where only two variables are set as the trainable to recover the source data from the received sig-
parameters in each iteration. Since the number nals. The weights of the model are optimized in a
of trainable parameters are comparable to that of supervised learning manner based on an end-to-
iterations, it can be easily trained within a shorter end loss on the recovery accuracy. By doing so,
period and with less training data compared to a the block structures in the conventional communi-
regular DNN while improving the performance of cation systems are no longer required. Moreover,
the orthogonal AMP detector in Rayleigh and cor- the end-to-end method has great potential to pro-
related MIMO channels. Therefore, this approach vide a universal solution for different channels.
can be scaled to massive MIMO communications As aforementioned, the weights of the DNN are
with great potential to be applied to time-varying usually learned based on the SGD with the gradi-
channels. ents of the loss function back-propagated from the
Apart from wireless signal compression and output layer to the input layer. Nevertheless, when
detection, DL has been exploited for various tasks the channel parameters are unknown in advance,
in physical layer communications. Compared with the gradients cannot back-propagate through the
traditional methods, it has shown higher robust- unknown channel since the gradients for updat-
ness to channels. For example, DNN has been ing the transmitter are blocked by the unknown
utilized in channel decoding and is more robust channel, which forestalls the learning of the end-to-
to variations of the additive white Gaussian noise end networks. The channel transfer function may
(AWGN) channel model [13]. In addition, DL can be pre-assumed to solve the issue, but any such
improve the system performance by exploiting the assumption would bias the learned models, repeat-
additional contextual information. For example, ing the pitfalls resulting from the likely discrepan-
in mmWave systems, DL can be used for beam cy between the assumed channel models and the
1 The data set and simulation
prediction, where some contextual information, actual channels. In addition, in real communication
codes can be downloaded
such as the locations of the receiver and the sur- systems, an accurate channel transfer function is from https://github.com/
rounding vehicles in vehicular networks, can be difficult to obtain in advance since the end-to-end haoyye/OFDM_DNN.

IEEE Wireless Communications • April 2019 97

Apart from wireless
Agent Environment
signal compression and
detection, DL has been
exploited for various Source data Transmitter Transmit signals Received Receiver Received
(State) (Action) signals data
tasks in physical layer Channel
communications. Com-
pared with traditional
methods, it has shown
higher robustness to End-to-end loss (Reward)

channels. For example, (a)

DNN has been utilized
in channel decoding Conditional GAN

and is more robust to Source

Transmitter Generator Receiver
Received
data data
variations of the addi-
tive white Gaussian
noise (AWGN) channel
Received Discriminator Received
model. pilot signals pilot signals
Transmit
signals

Received
Transmit Data Flow pilot signals
Channel

Gradients Flow

(b)

FIGURE 6. End-to-end communication system models: a) reinforcement learning based end-to-end com-
munication systems; b) conditional GAN based end-to-end communication systems.

channel often embraces different types of random consists of DNNs for the transmitter, the channel
effects, such as channel noise and time-varying, generator, and the receiver, which are trained iter-
which may be unknown or cannot be expressed atively. Since the conditional GAN learns to mimic
analytically. As shown in Fig. 6, we will introduce the channel effects, it acts as a surrogate channel
two methods to address the issue in this section. for the gradients to pass through, which enables
the training of the transmitter. The conditioning
Reinforcement Learning-based End-to-End Systems information for the conditional GAN is the transmit
In [15], a reinforcement learning based approach signals from the transmitter along with the received
has been proposed to circumvent the problem of pilot information used for estimating the channel.
missing gradients from channels when optimizing Therefore, the generated output distribution will
the transmitter. As shown in Fig. 6a, the transmit- be specific to the instantaneous channel and the
ter, converting the source data into the transmit transmit signals. As a result, the conditional GAN
symbols, is considered an agent while both the based end-to-end communication system can be
channel and the receiver are regarded as the envi- applied to more realistic time-varying channels. The
ronment. The agent will learn to take actions to simulation results in [4] confirm the effectiveness of
maximize the cumulative rewards emitted from the conditional GAN based end-to-end communi-
the environment. At each time, the transmit data cation system, by showing similar performance as
is regarded as the state observed by the trans- the Hamming (7,4) code with maximum-likelihood
mitter and the transmit signals are regarded as decoding (MLD).
the action taken by the transmitter. The end-to-
end loss on each sample will be calculated at the Conclusions and Future Directions
receiver and fed back to the transmitter as the We have demonstrated the great potential of DL
reward from the environment, which guides the in physical layer communications. By summarizing
training of the transmitter. By using the policy gra- how to apply DL in communication systems, the
dient algorithm, a standard reinforcement learning following research directions have been identified
approach, the transmitter can learn to maximize to bring intelligent physical layer communications
the reward, that is, optimize the end-to-end loss, from theory to practice.
without requiring the gradients from the channel. Can DL-based End-to-End Communications
Beat the Traditional? We have briefly introduced
Conditional GAN-based End-to-End Systems end-to-end communications. From the initial
In order to solve the missing gradient problem and research results in [4] and [15], the performance
lower the demands for the large amount of training of DL-based end-to-end communications is com-
data, a generative approach based on conditional parable with the traditional ones. However, it is
GAN has been proposed in [4]. As in Fig. 6b, the not clear whether DL-based end-to-end commu-
end-to-end learning of a communication system is nications will eventually outperform the traditional
enabled without requiring prior information of the ones in terms of performance and complexity or
channel by modelling the conditional distribution how much gain can be achieved. We are expect-
of the channel. In Fig. 6b, the end-to-end pipeline ing the answers to these questions soon.

98 IEEE Wireless Communications • April 2019

 Trade-Off between System Performance and [7] A. Mousavi and R. G. Baraniuk, “Learning to Invert: Signal The bloom of various
Recovery via Deep Convolutional Networks,” Proc. IEEE Int’l.
Training Efficiency: Existing work has shown the Conf. Acoustics Speech Signal Process. (ICASSP’17), New applications of learning
power of data-driven models in physical layer Orleans, LA, Mar. 2017, pp. 2272–76. techniques should be
communications. Even though a universal trans- [8] V. Papyan et al., “Theoretical Foundations of Deep Learning
largely credited to the
mitter/receiver can be optimized in the end-to- via Sparse Representations: A Multilayer Sparse Model and
Its Connection to Convolutional Neural Networks,” IEEE accessible real-world
end learning-based communication design, the Signal Process. Mag., vol. 35, no. 4, July 2018, pp. 72–89.
training process takes very long as all the commu- [9] C. A. Metzler, A. Mousavi, and R. G. Baraniuk, “Learned data sets, such as Ima-
nication blocks are merged. In order to improve D-AMP: Principled Neural Network Based Compressive geNet for computer
the training efficiency and achieve good system Image Recovery,” Advances Neural Info. Process. Systems
(NIPS), Long Beach, CA, Dec. 2017, pp.1772-1783.. vision. These open
performance, part of the communication blocks [10] C. K. Wen, W. T. Shih, and S. Jin, “Deep Learning for Mas- access data sets pro-
can be kept and model-driven DL methods can sive MIMO CSI Feedback,” IEEE Wireless Commun. Lett.,
be considered. Then we need to carefully design vol. 7, no. 5, Oct. 2018, pp. 748–51. vide an efficient way
the system to achieve a good trade-off between [11] N. Farsad, M. Rao, and A. Goldsmith, “Deep Learning for to compare the per-
Joint Source-Channel Coding of Text,” Proc. IEEE Int’l. Conf.
training efficiency and system performance. Acoustics Speech Signal Process. (ICASSP’18), Calgary, AB, formance of different
Communication Metric Learning: In tradi- Apr. 2018, pp. 2326–30. learning algorithms.
tional communication systems, the objective is [12] H. Ye, G. Y. Li, and B.-H. F. Juang, “Power of Deep Learn-
However, such a type
the error-free reconstruction of the transmit data, ing for Channel Estimation and Signal Detection in OFDM
Systems,” IEEE Wireless Commun. Lett., vol. 7, no. 1, Feb. of accessible data sets
while in real applications, the objective of sharing 2018, pp. 114–17.
the information may vary from task to task and [13] H. Kim et al., “Communication Algorithms via Deep Learn- for wireless communi-
the reconstruction metrics may not be satisfactory ing,” arXiv preprint arXiv:1805.09317, May 2018. cations are still under
for all tasks. For instance, bit-error rate is not a [14] S. Dörner et al., “Deep Learning Based Communication
over the Air,” IEEE J. Sel. Topics Signal Process., vol. 12, no. developed.
good metric for image and video transmission 1, Feb. 2018, pp. 132–43.
since it cannot reflect the properties of human [15] F. Aoudia and J. Hoydis. “End-to-End Learning of Commu-
visual perception. In end-to-end communication nications Systems Without a Channel Model,” arXiv preprint
systems, the metric should be revised to address arXiv:1804.02276, Dec. 2018.
specific requirements for each application. The
basic idea is that the transmit data will not be Biographies
Z hijin Q in ([email protected]) is a lecturer (assistant profes-
treated as equally important, the recovered data sor) at Queen Mary University of London. She was with Lan-
may contain transmission error, but the semantic caster University and Imperial College London as a lecturer
information contained in the data, which is further and research associate, respectively, from 2016 to 2018. Her
research interests include machine learning and compressive
employed for the application-specific tasks at the sensing in wireless communications and Internet of Things net-
receiver, should remain intact. works. She is an associate editor of IEEE Communications Let-
Open Access Real-World Data Sets: The ters. She received the Best Paper Award from IEEE GLOBECOM
bloom of various applications of learning tech- 2017, and the IEEE Signal Processing Society Young Author Best
Paper Award 2018.
niques should be largely credited to the acces-
sible real-world data sets, such as ImageNet for Hao Ye ([email protected]) received the B.E. degree in infor-
computer vision. These open access data sets pro- mation engineering from Southeast University, Nanjing, China,
vide an efficient way to compare the performance in 2012, and the M.E. in electrical engineering from Tsinghua
University, Beijing, China, in 2015. He is currently working
of different learning algorithms. However, such toward the Ph.D. degree in electrical engineering at the Georgia
a type of accessible data sets for wireless com- Institute of Technology, Atlanta, GA, USA. His research interests
munications are still under developed. The data include machine learning, signal detection, vehicular communi-
protection and privacy regulations further limit the cations, and radio resource management.
open access of real-world communication data. Geoffrey Ye Li ([email protected]) is a professor with Geor-
However, it is still essential to publish some data gia Tech. His general research is in signal processing and
sets, that is, channel responses under different machine learning for wireless communications. In these areas,
channel conditions, to speed up the development he has published over 500 articles with over 35,000 citations
and has been listed as a Highly-Cited Researcher by Thomson
of DL-based physical layer communications. Reuters. He has been an IEEE Fellow since 2006. He won the
IEEE ComSoc Stephen O. Rice Prize Paper Award and the
References Award for Advances in Communication, the IEEE VTS James
Evans Avant Garde Award and the Jack Neubauer Memorial
[1] T. O’Shea and J. Hoydis, “An Introduction to Deep Learning
for the Physical Layer,” IEEE Trans. Cogn. Commun. Netw., Award, the IEEE SPS Donald G. Fink Overview Paper Award,
vol. 3, no. 4, Dec. 2017, pp. 563–75. and the Distinguished ECE Faculty Achievement Award from
[2] H. He et al., “Model-Driven Deep Learning for Physical Layer Georgia Tech.
Communications,” arXiv preprint arXiv:1809.06059, Sept.
2018, to appear in IEEE Wireless Commun. Biing-Hwang (Fred) Juang is the Motorola Foundation Chair
[3] R. C. Gonzalez, “Deep Convolutional Neural Networks [lec- Professor and a Georgia Research Alliance Eminent Scholar at
ture notes],” IEEE Signal Processing Mag., vol. 35, no. 6, Nov. Georgia Tech. He received a Ph.D. degree from the University
2018, pp. 79–87. of California, Santa Barbara and has conducted research at Bell
[4] H. Ye et al., “Channel Agnostic End-to-End Learning Based Labs, including serving as Director of Acoustics and Speech
Communication Systems with Conditional GAN,” Proc. IEEE Research. He joined Georgia Tech in 2002. Prof. Juang’s aca-
Global Commun. Conf. Workshops (GC Wkshps’18), Abu demic distinctions include: several best paper awards, IEEE
Dhabi, UAE, Dec. 2018, pp. 1-5.. Fellow, the Technical Achievement Award from the IEEE SPS
[5] K. Arulkumaran et al., “Deep Reinforcement Learning: A for contributions to the field of speech processing and com-
Brief Survey,” IEEE Signal Process. Mag., vol. 34, no. 6, Nov. munications, Bell Labs Fellow, member of the U.S. National
2017, pp. 26–38. Academy of Engineering, Academician of Academia Sinica,
[6] Z. Qin et al., “Sparse Representation for Wireless Commu- the IEEE J.L. Flanagan Field Medal Award in Audio, Speech
nications: A Compressive Sensing Approach,” IEEE Signal and Acoustics, and a Charter Fellow of the National Academy
Process. Mag., vol. 35, no. 3, May 2018, pp. 40–58. of Inventors.

IEEE Wireless Communications • April 2019 99