You are on page 1of 4

1240 IEEE COMMUNICATIONS LETTERS, VOL. 23, NO.

7, JULY 2019

CNN-Based Precoder and Combiner Design in mmWave MIMO Systems


Ahmet M. Elbir
Abstract— Hybrid beamformer design is a crucial stage in channel state information (CSI) feedback problem. In [14],
millimeter-wave (mmWave) MIMO systems. In this letter, we pro- a sub-optimum method is proposed based on the support vector
pose a convolutional neural network (CNN) framework for the machines (SVMs) for analog beamforming vector selection.
joint design of precoder and combiners. The proposed network
accepts the input of channel matrix and gives the output of analog Very recently, a DL based hybrid beamforming is considered
and baseband beamformers. Previous works are usually based in [18] where only precoder design is considered whereas
on the knowledge of steering vectors of array responses which is joint precoder and combiner design is used in massive MIMO
not always accurately available in practice. The proposed CNN system where the beamforming is required in both end of
framework does not require such a knowledge, and it provides the communication [6]. The proposed network architecture
higher performance in capacity compared with the conventional
greedy- and optimization-based algorithms. in [18] is based on multi-layer perceptrons which do not
Index Terms— mmWave, MIMO, hybrid beamforming, deep effectively extract the hidden features inherit in the input
learning, convolutional neural network. data [9], [13]. In order to achieve feature extraction and obtain
better performance, we propose a convolutional neural network
I. I NTRODUCTION (CNN) framework for mmWave massive MIMO systems.

H YBRID beamforming is a promising architecture to be


used in next generation millimeter wave (mmWave)
MIMO (Multiple Input Multiple Output) systems where
In this study, we propose a CNN-based framework with
two CNNs, each of which is dedicated to estimate the analog
precoders and combiners respectively. The CNNs accept the
robust beamforming performance is provided with smaller input of channel matrix and give the beamformer weights at
cost and less number of fully-digital beamformers [1]–[4]. the output. In order to train the network we generated different
Several methods are proposed to design the hybrid channel realizations with synthetic noise added to each input
beamformers [3]–[7]. In [6], a greedy-based approach, data. A decoupled optimization problem is formulated and
orthogonal matching pursuit (OMP), is proposed where the solved to obtain the best beamformers providing the highest
analog precoder and combiners are selected from a dictio- spectral efficiency. Using the best analog beamformers, we
nary of transmit and receive array responses. This algorithm construct the input-output pairs of the network. Since the
requires the knowledge of the user direction-of-arrival/aperture best beamformers are the optimum solution of the problem,
(DOA/DOD) angles to construct such a dictionary. Using the proposed CNN framework enjoys better spectral efficiency
the connection between the optimum and the hybrid beam- and less computation time. Furthermore, our CNN approach
formers, [7] proposes an alternating minimization approach to does not require the knowledge of array responses of users’
estimate the analog and baseband beamformers based on phase DOA/DOD angles which are not always accurately available
extraction. in practical scenarios.
The above works provides optimization-based and greedy-
based solutions for hybrid beamforming problem. However
II. S IGNAL M ODEL AND P ROBLEM F ORMULATION
achieving the optimum solution and the computation time
are the main drawbacks of the above techniques. In order to In this work, we consider a single user mmWave MIMO
circumvent this issue, we consider deep learning (DL)-based communication system with multiple antennas. Let NS be
techniques for the hybrid beamforming problem [8]. DL the number of data streams to be transmitted from the base
has several advantages such as low computational complex- station (BS) with NT transmit antennas to the user with NR
ity when solving optimization-based or combinatorial/greedy antennas. The BS is equipped with NTRF analog phase shifters
RF
search problems and the ability to extrapolate new features with analog beamformer FRF ∈ CNT ×NT and baseband
RF
from a limited set of features contained in a training set [9]. beamformer FBB ∈ CNT ×NS . Then the transmitted signal
A great deal of attention is received for DL-based tech- becomes x = FRF FBB s where s ∈ CNS is the symbol
niques in communications society for the problems such as vector desired to be transmitted and E{ssH } = INS /NS .
channel estimation [10]–[12] DOA estimation [12], antenna The analog beamformers are unitary matrices with equal-norm
selection [13], and analog beam selection [14]. An end-to- elements, i.e., [[FRF ]:,i [FRF ]H
:,i ]i,i = 1/NT and we have power
end communication scenario is modeled in [15] and [16] by constraint on the transmitter as ||FRF FBB ||F = NS . We can
using auto-encoders where single-input-single-output (SISO) write the received signal at the NR antennas for a narrowband
systems are considered. [17] also uses auto-encoders for the block-fading channel as

Manuscript received April 22, 2019; accepted May 6, 2019. Date of y = ρHFRF FBB s + n, (1)
publication May 9, 2019; date of current version July 10, 2019. The associate
editor coordinating the review of this letter and approving it for publication where y ∈ CNR and ρ is the average received power. n ∈ CNR
was J. Choi. denotes the additive white Gaussian noise (AWGN) with
The author is with the Department of Electrical and Electronics Engineering,
Düzce University, Duzce 81620, Turkey (e-mail: ahmetmelbir@gmail.com). n ∼ CN (0, σn2 INR ) and H ∈ CNR ×NT is the channel
Digital Object Identifier 10.1109/LCOMM.2019.2915977 matrix with E{||H||F } = NR NT . In mmWave transmission,
1558-2558 © 2019 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
ELBIR: CNN-BASED PRECODER AND COMBINER DESIGN IN mmWAVE MIMO SYSTEMS 1241

the channel can be represented by Saleh-Valenzuela (SV) where FRF and WRF denote the feasible sets of analog
model [19] where the clustered channel model is used as the beamformers which obey the constraints defined for FRF
contribution of Nc clusters of Nray paths as and WRF . Obtaining the real-time solution to the prob-
lem in (4) is impractical due to the complexity of several
Nray
Nc 
 (ij) (ij) (ij) (ij) matrix variables. To cast the problem in (4) more effec-
H=γ αij gR (ΘR )gT (ΘT )aR (ΘR )aH
T (ΘT ), tively, we first define the sets FRF and WRF . Note that
i=1 j=1
the analog beamformers FRF , WRF are related with the
(ij) (ij)
where ΘR
(ij) (ij)
= (φR , θR ) and ΘT
(ij) (ij)
= (φT , θT )
(ij) (ij) array responses aT (ΘT ), aR (ΘR ) through linear trans-
respectively denote the angle of arrivals and angle of formation [6]. Hence the feasible RF beamformer sets can
(1) (Q ) (q )
departures. We denote the angular parameters φ and θ be formed as FRF = {FRF , . . . , FRFF } where FRFF =
(ij)
as the aT (ΘT ), i = 1, . . . ,Nc , j = 1, . . . , Nray for qF =
 azimuth and the elevation angles respectively. N
γ = NT NR /(Nc Nray ) is the normalization factor and αij 1, . . . , QF . QF = N path RF is the number RF precoder candi-
T
is the complex channel gain associated with the ith scattering dates and Npath = Nc Nray . The feasible set for RF combiner
cluster and jth path for i = 1, . . . , Nc and j = 1, . . . , Nray . (1)
is similarly defined as WRF = {WRF , . . . , WRFW } where
(Q )
(ij) (ij)
gR (ΘR ) and gT (ΘT ) are the antenna element gains for (qW )
WRF = aR (Θ
(ij)
(ij)  R ), i = 1, . . . , Nc , j = 1, . . . , Nray and
receive and transmit antennas respectively. aR (ΘR ) and Npath
(ij) QW = N RF . Now we can present the joint precoder and
aT (ΘT ) are NR ×1 and NT ×1 steering vectors representing R
combiner design problem as follows
the array responses at the receiver and transmitters
respectively. The nth element of the steering vector  ρ −1 H
q̄F , q̄W = argmax log2 INS + Λ WBB WH
(ij) (ij)
aR (ΘR ) is given as [aR (ΘR )]n = exp{− 2π T (ij)
λ pn r(ΘR )},
qF ,qW NS n RF

T  
where pn = [xn , yn , zn ] is the position of the nth receive × HFRF FBB FH H H
BB FRF H WRF WBB ,

(ij)
antenna in Cartesian coordinate system and r(ΘR ) = (q ) (q )
(ij) (ij) (ij) (ij) (ij) T s.t.: FRF = FRFF , WRF = WRFW ,
[sin(φR ) cos(θR ), sin(φR ) sin(θR ), cos(θR )] . The
(ij)
transmit side steering vector aT (ΘT ) can be defined in a FBB = (FH
RF FRF )
−1 H
FRF Fopt ,
(ij)
similar way as for aR (ΘR ). In order to generate the labels WBB = (WH RF ΛWRF )
−1
(WH RF ΛW ), (5)
opt

(precoder and combiners) the proposed CNN frameworks


requires the perfect CSI. However, we use imperfect channel where q̄F , q̄W represent the selected elements in the feasible
matrices both in the training and testing stages. sets. Λ is the covariance of the array output in (1) which is
The transmitted signal is received and processed by analog given by Λ = NρS HFRF FBB FH H H 2
BB FRF H + σn INRS . F
opt
and
opt
and baseband combiners as ỹ = WH H
BB WRF y, i.e.,
W are the optimum baseband beamformers which can be
√ obtained from the singular value decomposition (SVD) of the
ỹ = ρWH H
BB WRF HFRF FBB s + WBB WRF n,
H H
(2) channel matrix. Let U ∈ CNR ×rank(H) and V ∈ CNT ×rank(H) be
RF the left and the right singular value matrices of H respectively,
where WRF ∈ CNR ×NR is the analog combiner with where the SVD of H ∈ CNR ×NT is H = UΣVH where Σ
the constrained [[WRF ]:,i [WRF ]H
:,i ]i,i = 1/NR and WBB ∈ is rank(H) × rank(H) matrix composed of the singular values
RF
NR ×NS
C denotes the baseband combiner matrix. By assum- of H in descending order. By decomposing Σ and V as Σ =
ing that the Gaussian symbols are transmitted through the diag{Σ(1) , Σ(2) }, V = [V(1) , V(2) ] where V(1) ∈ CNT ×NS
mmWave channel, we can define the spectral efficiency and V(2) ∈ CNT ×NR −NS respectively, one can readily select
achieved by the hybrid beamforming [3]–[6] as the unconstrained precoder as Fopt = V(1) [6]. Using the
 ρ −1 H unconstrained beamformer Fopt , Wopt can be computed as [20]
RHYB = log2 INS + Λ WBB WH RF H
NS n  1  optH H opt NS σn2 −1 optH H H

× FRF FBB FH H H  Wopt = H HF + .
BB FRF H WRF WBB , (3) F I NS F H
ρ ρ
where Λ= 2 H H
n σn WBB WRF WRF WBB ∈ C
NS ×NS
is the covari- The solution of (5) requires to visit QF QW nodes which is
ance matrix of the noise term in (2) after combining. computationally prohibitive. In order to reduce the complexity,
Hence the aim in this work is to estimate the hybrid (5) is decomposed into two different problems where precoders
beamformers FRF , WRF , FBB and WRF that maximize the (FRF and FBB ) and combiners (WRF and WBB ) are sepa-
spectral efficiency as in (3) given the channel matrix H. rately estimated. By doing so, the complexity is reduced from
QF QW to QF + QW . In order to find the precoders we solve
III. H YBRID B EAMFORMER D ESIGN
the following problems, i.e.,
The optimization problem for joint estimation of hybrid  ρ H H
beamformers F̂RF , F̂BB , ŴRF , ŴBB can be stated as follows q̄F = argmax log2 |INS + 2
(Wopt Wopt )−1 Wopt
qF NS σn

argmax RHYB × HFRF FBB FH H H
BB RF H W | ,
F opt
FRF ,FBB ,WRF ,WBB
(q )
s.t.: FRF ∈ FRF , ||FRF FBB ||2F = NS , s.t.: FRF = FRFF , FBB = (FH
RF FRF )
−1 H
FRF Fopt ,
WRF ∈ WRF , (4) (6)
1242 IEEE COMMUNICATIONS LETTERS, VOL. 23, NO. 7, JULY 2019

Algorithm 1 Training Data Generation


Input: L, N , NT , NR , NTRF , NRRF , SNRTRAIN .
Output: Training data DF and DW .
(n) N (n) N
1: Generate {H(n) }N n=1 with {FRF }n=1 and {WRF }n=1 .
2: for 1 ≤ n ≤ N and 1 ≤ l ≤ L do
3: [H(l,n) ]i,j ∼ CN ([H(l) ]i,j , σTRAIN
2
).
(qF ,l,n)
4: Find q̄F by solving (6) for FRF , 1 ≤ qF ≤ QF .
(l,n) (l,n) (q̄ ,l,n)
5: Construct F̂RF and FBB from FRFF .
(qW ,l,n)
6: Find q̄W by solving (7) for WRF , 1 ≤ qW ≤ QW .
Fig. 1. The proposed CNN framework for precoder (CNNF at the top) and (l,n) (l,n) (q̄ ,l,n)
combiner design (CNNW at the bottom). 7: Construct ŴRF and WBB from WRFW .
(l,n)
 8: [[X ]:,:,1 ]i,j = |[H(l,n) ]i,j |.
ρ
q̄W = argmax log2 |INS + 2
(WH H
BB WRF WRF WBB 9: [[X
(l,n)
]:,:,2 ]i,j = Re{[H(l,n) ]i,j } .
qW N S σn (l,n)
H  10: [[X ]:,:,3 ]i,j = Im{[H(l,n) ]i,j } ∀ij.
× WH H
RF WBB )
−1
HFopt Fopt HH WRF WBB | , 11:
(l,n) (l,n) (l,n)
zF = ∠vec{F̂RF }, zW = ∠vec{ŴRF }.
(l,n)

(q )
s.t.: WRF = WRFW , 12: end for n, l
13: Training data for CNNF and CNNW is obtained as
WBB = (WH
RF ΛWRF )
−1
(WH RF ΛW ),
opt
(1,1) (L,N )
ρ H DF = ((X(1,1) , zF ), . . . , (X(L,N ) , zF )),
Λ= HFopt Fopt HH + σn2 INR . (7) (1,1) (L,N )
DW = ((X(1,1) , zW ), . . . , (X(L,N ) , zW )).
NS
Once (6) and (7) are solved, the analog beamformers are
(q̄ ) (q̄ )
constructed as F̂RF = FRFF and ŴRF = WRFW . The
baseband beamformers can also be obtained accordingly. 3 × 3N L. In order to obtain the output data the problems in
(6) and (7) are solved ∀n, l. Then the output data of each
network is obtained. We summarize the algorithmic steps of
IV. CNN-BASED A PPROACH
the training data generation in Algorithm 1.
In this section, we present our CNN framework for joint
precoder and combiner design which is shown in Fig. 1. V. N UMERICAL S IMULATIONS
The proposed network is composed of two CNNs with 8 layers In this section, we evaluate the performance of our CNN
which have identical structures except the last layer. The first framework (referred to as HBDL, Hybrid Beamforming via
layer is the input layer of size NR × NT × 3 with c = 3 Deep Learning) and compare it with the state-of-the-art tech-
channels. The first channel of the input is the element-wise niques such as SOMP [6] and PE-Alt-Min [7]. Uniform square
absolute value of the channel matrix as [[X]:,:,1 ]i,j = |[H]i,j |. arrays are considered with half wavelength spacing with NR =
The second and the third channels are defined as the real and NT = 36 antennas. The number of analog beamformers are
the imaginary parts of the channel matrix as [[X]:,:,2 ]i,j = NRRF = NTRF = 4. The feasible sets FRF , WRF are used
Re{[H]i,j } and [[X]:,:,3 ]i,j = Im{[H]i,j }. The second and third for training only, and the output from CNN can be directly
layer are the convolutional layers with 32 filters of size 2 × 2. used for analog beamforming since the analog beamformer
The fourth and sixth layers are fully connected layers with does not have to lie in the set of array response vectors.
1024 units. There are dropout layers after each fully connected The CNNs are fed with the training data generated for N =
layers (the fifth and seventh layers) with %50 probability. L = 100. For each channel matrix realization, the propagation
The output layer of CNNF is of size NT NTRF × 1 which is environment is modeled with Nc = 4 and Nray = 5 for each
2
the vectorized version of the phases of FRF . Similarly, the clusters with σΘ = 5◦ for all transmit and receive azimuth
size of the output layer of CNNW is NR NRRF × 1. The and elevation angles which are uniform randomly selected
complexity of a CNN is directly proportional with the number from the interval [−60◦, 60◦ ] and [−20◦, 20◦ ] respectively.
of parameters which, in our case, calculated as C 2 (2Ncv (wh+ The proposed network is realized in MATLAB on a PC
50
1) + 2(Nf c + 1) · 100 ) [21]. Here C = 3 is the number of with 768-core GPU. Stochastic gradient decent algorithm is
channels, w = h = 2 is the filter size, Ncv = 32 is the used to update the network parameters with the learning rate
number of filters, Nf c = 1024 is the number of units in the 0.005 and mini-batch size 500 for 200 epochs. As a loss
fully connected layer for %50 dropout probability. Hence the function, we use the negative log-likelihood or cross-entropy
CNN structure in Fig. 1 has 12105 parameters. loss [9]. In the training process, 70% and 30% of all data
In data generation, N different realizations of channel matri- generated are selected as the training and validation datasets,
ces H(n) for different user locations are generated together respectively. Validation aids in hyperparameter tuning during
(n) (n)
with the corresponding sets FRF and WRF . Then for each the training phase to avoid the network simply memorizing the
realization, L noisy channel matrices are obtained where the training data rather than learning general features for accurate
added element-wise synthetic noise is defined by SNRTRAIN = prediction with new data. The validation data is used to test the
|[H] |2
20 log10 ( σ2 i,j ). To account for the changes in the wireless performance of the network in the simulations for JT = 100
TRAIN
environment, we use three different SNRTRAIN levels. Hence Monte Carlo trials. In order to prevent the similarity between
the total size of the training input data becomes NR × NT × the test data and the training data we also add synthetic
ELBIR: CNN-BASED PRECODER AND COMBINER DESIGN IN mmWAVE MIMO SYSTEMS 1243

Fig. 2. Spectral efficiency versus SNR for (a) NR = NT = 25, NS = 1; (b) NR = NT = 36, NS = 2; (c) NR = NT = 36, NS = 3.

noise to the test data where the SNR in testing is defined [4] A. Alkhateeb, G. Leus, and R. W. Heath, Jr., “Limited feedback hybrid
|[H] |2 precoding for multi-user millimeter wave systems,” IEEE Trans. Wireless
similar to SNRTRAIN as SNRTEST = 20 log10 ( σ2i,j ) and Commun., vol. 14, no. 11, pp. 6481–6494, Nov. 2015.
TEST
SNRTRAIN ∈ {10, 15, 20}dB is selected. [5] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Hybrid
In Fig. 2, the spectral efficiency for different algorithms is precoding for millimeter wave cellular systems with partial channel
knowledge,” in Proc. Inf. Theory Appl. Workshop (ITA), Feb. 2013,
presented for NS = {1, 2, 3} and SNRTEST = 10dB. As it pp. 1–5.
is seen, HBDL provides better performance as compared to [6] O. El Ayach, S. Rajagopal, S. Abu-Surra, Z. Pi, and R. W. Heath, Jr.,
the optimization-based method PE-Alt-Min and greedy-based “Spatially sparse precoding in millimeter wave MIMO systems,” IEEE
Trans. Wireless Commun., vol. 13, no. 3, pp. 1499–1513, Mar. 2014.
algorithm SOMP. The performance plot “Best” denotes the [7] X. Yu, J.-C. Shen, J. Zhang, and K. B. Letaief, “Alternating minimization
performance of the test data without prediction. We observe algorithms for hybrid precoding in millimeter wave MIMO systems,”
that HBDL is very close to the best performance as well IEEE J. Sel. Topics Signal Process., vol. 10, no. 3, pp. 485–500,
Apr. 2016.
as the fully-digital beamformer. HBDL effectively selects the [8] D. Yu and L. Deng, “Deep learning and its applications to signal and
analog beamformers from the feasible sets which maximizes information processing [exploratory DSP],” IEEE Signal Process. Mag.,
the spectral efficiency. The effectiveness of HBDL is attributed vol. 28, no. 1, pp. 145–154, Jan. 2011.
[9] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521,
to the best selection of analog beamformers which are the pp. 436–444, May 2015.
optimum solution of (4) through the SVD of the channel [10] A. Alkhateeb, S. Alex, P. Varkey, Y. Li, Q. Qu, and D. Tujkovic,
matrix [6]. SOMP has poor performance due the fact that it “Deep learning coordinated beamforming for highly-mobile millime-
ter wave systems,” Apr. 2018, arXiv:1804.10334. [Online]. Available:
cannot select the “best” set of array responses from the dic- https://arxiv.org/abs/1804.10334
tionary. While PE-Alt-Min has sufficiently good performance, [11] H. Ye, G. Y. Li, and B.-H. Juang, “Power of deep learning for channel
HBDL performs better even when the output of PE-Alt-Min estimation and signal detection in OFDM systems,” IEEE Wireless
Commun. Lett., vol. 7, no. 1, pp. 114–117, Feb. 2018.
is inserted to the feasible sets used for HBDL. [12] H. Huang, J. Yang, H. Huang, Y. Song, and G. Gui, “Deep learn-
To compare the computation time of the algorithms we ing for super-resolution channel estimation and doa estimation based
consider the same settings and observe that HBDL spends massive MIMO system,” IEEE Trans. Veh. Technol., vol. 67, no. 9,
pp. 8549–8560, Sep. 2018.
about 0.020s to compute both precoder and combiners whereas [13] A. M. Elbir, K. V. Mishra, and Y. C. Eldar, “Cognitive radar antenna
SOMP and PE-Alt-Min take about 0.450s and 1.200s respec- selection via deep learning,” IET Radar, Sonar Navigat., to be published.
tively. [14] Y. Long, Z. Chen, J. Fang, and C. Tellambura, “Data-driven-based analog
beam selection for hybrid beamforming under mm-wave channels,”
IEEE J. Sel. Topics Signal Process., vol. 12, no. 2, pp. 340–352,
VI. C ONCLUSIONS May 2018.
In this work, a CNN framework is proposed for the joint [15] S. Dörner, S. Cammerer, J. Hoydis, and S. T. Brink, “Deep learning
based communication over the air,” IEEE J. Sel. Topics Signal Process.,
estimation of precoder and combiners in hybrid beamform- vol. 12, no. 1, pp. 132–143, Feb. 2018.
ing problem. We show that the proposed network archi- [16] V. Raj and S. Kalyani, “Backpropagating through the air: Deep learning
tecture provides better spectral efficiency as compared to at physical layer without channel models,” IEEE Commun. Lett., vol. 22,
no. 11, pp. 2278–2281, Nov. 2018.
the optimization-based and greedy-based algorithm. In future [17] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO
work, we reserve the case when the training data is small CSI feedback,” IEEE Wireless Commun. Lett., vol. 7, no. 5, pp. 748–751,
where transfer learning-like approaches can be developed. Oct. 2018.
[18] H. Huang, Y. Song, J. Yang, G. Gui, and F. Adachi, “Deep-learning-
R EFERENCES based millimeter-wave massive MIMO for hybrid precoding,” IEEE
Trans. Veh. Technol., vol. 68, no. 3, pp. 3027–3032, Mar. 2019.
[1] J. G. Andrews et al., “What will 5G be?” IEEE J. Sel. Areas Commun., [19] R. Méndez-Rial, C. Rusu, A. Alkhateeb, N. Gonzalez-Prelcicy, and
vol. 32, no. 6, pp. 1065–1082, Jun. 2014. R. W. Heath, Jr., “Channel estimation and hybrid combining for
[2] F. Rusek et al., “Scaling up MIMO: Opportunities and challenges with mmWave: Phase shifters or switches?” in Proc. IEEE Inf. Theory Appl.
very large arrays,” IEEE Signal Process. Mag., vol. 30, no. 1, pp. 40–60, Workshop, Feb. 2015, pp. 90–97.
Jan. 2013. [20] T. Kailath, B. Hassibi, and A. H. Sayed, Linear Estimation.
[3] A. Alkhateeb, O. El Ayach, G. Leus, and R. W. Heath, Jr., “Channel Upper Saddle River, NJ, USA: Prentice-Hall, 2000.
estimation and hybrid precoding for millimeter wave cellular systems,” [21] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
IEEE J. Sel. Topics Signal Process., vol. 8, no. 5, pp. 831–846, large-scale image recognition,” Sep. 2014, arXiv:1409.1556. [Online].
Oct. 2014. Available: https://arxiv.org/abs/1409.1556

You might also like