This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 1

DS-NLCsiNet: Exploiting Non-local Neural

Networks for Massive MIMO CSI Feedback
Xiaotong Yu, Xiangyi Li, Huaming Wu, Member, IEEE and Yang Bai

Abstract—Channel state information (CSI) feedback plays an Recently, deep learning (DL)-based methods have been
important part in frequency division duplex (FDD) massive introduced to CSI feedback tasks, and have shown great
multiple-input multiple-output (MIMO) systems. However, it is potential in CSI recovery [6]. Compared to CS-based meth-
still facing many challenges, e.g., excessive feedback overhead,
low feedback accuracy and a large number of training pa- ods, DL-based methods achieve a significant improvement in
rameters. In this paper, to address these practical concerns, model performance and computational speed. Wen et al. [7]
we propose a deep learning (DL)-based CSI feedback scheme, proposed an autoencoder (AE)-based network named CsiNet,
named DS-NLCsiNet. By taking advantage of non-local blocks, which uses an encoder to compress the channel matrices into
DS-NLCsiNet can capture long-range dependencies efficiently. codewords, and a decoder to transform the codewords into re-
In addition, dense connectivity is adopted to strengthen the
feature refinement module. Simulation results demonstrate that covered channel matrices. Exploiting convolutional operation
DS-NLCsiNet achieves higher CSI feedback accuracy and better and ResNet architecture [8] in the CSI feedback tasks, CsiNet
reconstruction quality for the same compression ratio, when outperforms existing CS-based algorithms at all compression
compared to state-of-the-art compression schemes. ratios. Recurrent neural network (RNN) is widely utilized in
Index Terms—Massive MIMO, Frequency Division Duplex new CSI feedback frameworks, such as CsiNet-LSTM [9],
(FDD), CSI Feedback, Non-local Neural Networks, Densely RecCsiNet [10] and ConvlstmCsiNet [11]. CsiNet-LSTM [9]
Connected Convolutional Networks. focused on modifying the channel recovery module, while
RecCsiNet [10] enhanced the feature compression and decom-
I. I NTRODUCTION pression module, while considering the temporal correlation of
the channel matrices to further improve the recovery quality.
R ECENTLY, massive multiple-input multiple-output
(MIMO) has emerged as one of the pivotal technologies
for fifth-generation (5G) wireless communication systems [1].
ConvlstmCsiNet [11] further exploited depthwise separable
convolutions to reduce the computational complexity and
Equipped with multiple transmitters and receivers, massive adopted convLSTM and Pseudo-3D to explore the spatial-
MIMO systems have achieved great efficiency in terms of temporal representation of channel information. In addition,
system capacity and anti-interference ability. In frequency CRNet [12] was based on an inception model to adapt to
division duplex (FDD) systems, one of the key procedures changes in granularity, and proposed an advanced training
is to exploit CSI at the base station (BS), which is essential scheme to enhance the network performance. CsiNetPlus [13]
for performance improvement. In conventional FDD MIMO has investigated the influence of convolutional kernels. CoC-
systems, the downlink CSI is obtained at the user equipment siNet [14] utilized the correlation between nearby UEs to
(UE), and then fed back to the BS through feedback links recover CSI cooperatively. CS-ReNet [15] can significantly
without compression. However, this method is prohibited in reduce the feedback overhead and lower the complexity of im-
massive MIMO systems and it is difficult to acquire a large plementing CS at the UE. Guo et al. [16] discussed the trend of
amount of accurate CSI in practical FDD systems since the computational complexity of neural networks and introduced
feedback overhead is extremely huge [2]. compression and acceleration techniques for communication
The challenge of CSI feedback in massive MIMO systems systems.
has stimulated numerous studies. Currently, compressive sens- To significantly boost the correctness of CSI feedback and
ing (CS)-based methods [3]–[5] are exploited to compress reduce the computational complexity of neural networks, we
the CSI, but these methods are not effective enough since design a novel CSI feedback architecture based on non-local
they require strict sparsity, so that channel matrices can neural networks [17], where non-local blocks are applied to ex-
hardly fit. Moreover, although CS simplifies the compression tract long-distance dependencies. In addition, we improve the
process, it turns decompression into an optimization problem, RefineNet module with dense connectivity [18] to strengthen
which requires an iterative algorithm and thus increases huge the feature propagation and enhance the information flow.
computational and time costs. The scheme brings improvements in terms of CSI feedback
accuracy and reconstruction quality.
X. Yu and Y. Bai are with the School of Mathematics, Tianjin University, The main contributions of this paper are listed as follows:
Tianjin 300072, China (e-mail: {xiaotongyu, baiyang123}@tju.edu.cn).
X. Li and H. Wu are with the Center of Applied Mathematics, Tianjin Uni- • We propose an innovative DL-based CSI feedback and
versity, Tianjin 300072, China (e-mail: {xiangyi li, whming}@tju.edu.cn). recovery mechanism, referred to as DS-NLCsiNet, which
This work is supported by the National Natural Science Founda- has the potential for practical deployment on real FDD
tion of China (61801325), the Huawei Innovation Research Program
(HO2018085138) and the Natural Science Foundation of Tianjin City (18JC- MIMO systems.
QNJC00600) (Corresponding author: Huaming Wu) • In DS-NLCsiNet, non-local blocks from non-local neural

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 2

Fig. 1: The architecture of proposed DS-NLCsiNet.

networks are applied to modify the feature extraction After performing the 2D-DFT and truncation operation, we
module and improve its efficiency in capturing long- separate the channel matrix H into real and imaginary parts.
range dependencies. Furthermore, dense connectivity is Then we feed it into the autoencoder network as depicted in
utilized to significantly enhance the recovery quality by Fig. 1, which includes the encoder and decoder. The encoding
encouraging feature reuse. and decoding procedures of CSI can be expressed as follows,
• Experimental results show DS-NLCsiNet can recover CSI respectively.
more accurately and improve the quality of recovered CS
t = fen (H), (4)
significantly when compared with some existing methods.
b = fde (t),
H (5)
II. S YSTEM M ODEL where the encoder compresses the CSI matrix of size q into
We consider a single-cell downlink FDD massive MIMO a codeword t of length p, and then sends it back to the
system constituted of Nt (Nt  1) antennas at the BS as well BS for CSI recovery. Then, the compression ratio (CR) is
as a single antenna at each user equipment (UE). The sys- defined as CR = p/q. After the BS finishes decompressing
tem is operated in orthogonal frequency-division multiplexing the codeword t to the original channel, we can obtain the
(OFDM) with N ec subcarriers. recovered channel matrix H
b by performing zero filling and
The received signal at the nth subcarrier is given as: inverse DFT procedure.
H
yn = h
en vn xn + zn , (1)
III. DS-NLC SI N ET
H
where hen ∈ CNt ×1, vn ∈ CNt ×1 , xn ∈ C and zn ∈ C denote The architecture of the proposed DS-NLCsiNet is shown in
the channel vector in the frequency domain, precoding vector Fig. 1, constituted of an encoder at the UE and a decoder at
designed by the BS, modulated transmit data symbol, and ad- the BS.
ditive Gaussian white noise at the nth subcarrier, respectively. In DS-NLCsiNet, it receives the truncated matrix H of size
Then the downlink CSI matrix H e is firstly obtained at the UE Nt × Nc × 2 as input and sends it to a 1 × 1 convolution
side via the downlink pilots, which can be modeled as: for initial information interaction. Then a non-local block
is applied to extract features, especially for capturing long-
e = [h
H e1 , h e e ] ∈ CNt ×Nec .
e2 , · · · , h (2)
Nc distance dependencies on the structure of channel matrix. The
The total number of feedback elements is Nt × N ec , which output of the non-local block remains the same shape with
will lead to high feedback overhead that beyond the system the input, and is then fed to the reshaping layer to stretch
capacity in a massive MIMO system. In order not to concen- into a 2Nc Nt -sized vector. The dense layer compresses the
trate on complicated details and challenges, we hypothesize vector into the p-sized (p < 2Nc Nt ) real-valued codeword
that perfect CSI has been acquired by the UE, and the BS t, where p satisfies the compression ratio (CR) standard:
can process the precoding vector vn as long as it receives the CR = p/2Nc Nt . The codeword is then fed back to the BS.
downlink CSI feedback H. e After the decoder at the BS receives the codeword, it first
To reduce feedback overhead, H e can be further transformed decompresses the p-sized codeword and reshapes it into a Nt ×
into a sparsified matrix H̄ in the angular-delay domain via a Nc × 2 sized rough recovery of H. Further refinements are
two-dimensional discrete Fourier transform (2D-DFT) opera- divided into two parts: Global Structure Refinement and Local
tion as follows: Detailed Refinement. A non-local block is first deployed to
e H
H̄ = Fd HF help reconstruct the global structure of CSI matrix, utilizing
a , (3)
its high efficiency of transferring information between remote
where Fd ∈ CNc ×Nc and Fa ∈ CNt ×Nt are both DFT matrices pixels. Then two DS-RefineNet blocks are used to supplement
e e

[9]. Moreover, since the time delay between multipath arrivals the local details, where the convolutions are local operations,
lies within a limited time period, only the first few columns which are more suitable for detailed reconstruction. Following
of H̄ having distinct non-zero values [7]. Thus, we only retain the DS-RefineNets, a 3×3 convolutional layer is implemented
the first Nc (Nc < Nec ) columns, and remove the rest columns. to scale the values to the [0, 1]. The final reconstruction of H is
H̄ is then truncated to a Nt × Nc sized CSI matrix H, which generated. For each convolutional layer, we use leaky ReLU as
still requires huge overhead for the massive MIMO system. the activation function and place a batch normalization layer.

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 3

A. Non-local Block local block.jpg

Existing DL-based CSI feedback architectures usually ex-
ploit convolutional or recurrent operations to extract the fea-
tures of the channel matrix. However, these two methods
can only deal with one local neighborhood at a time. To
obtain a larger resolution view, we usually need to repeat
these operations, which is computationally inefficient and will
cause optimization difficulties. Therefore, we introduce non-
local (NL) blocks from non-local neural networks [17] to
the CSI Feedback architecture, in order to capture long-range
dependencies well.
The NL-block is specially designed for sequence data
(spatial, temporal or spatial-temporal) and can directly pass
information between any two positions. The main idea comes
from the NL-Mean algorithm for image denoising, that is,
displaying mean operation on all image blocks, which is
calculated by:
Fig. 2: The structure of a non-local block with embedded Gaussian
version. “×” denotes the channel matrix multiplication, and “+”
X
u(xi ) = w(xi , y)v(y). (6)
denotes the element-wise sum. The Gaussian version can be imple-
y∈Ω
mented by removing θ and ∅. For convenience, the input size is set
To highlight commonalities and eliminate differences (usually Nc = Nt = 32 in our model.
noise), the normalized weight coefficient w(xi , y) is involved
here, i.e., the more similar block y with the output u(xi ), the
higher weight is given. turn from the original shape 1024 × 1024 to the downsized
Similar to the NL-Mean operation as shown in Eq. 6, the shape 256 × 256. Meanwhile, the number of channels is
generic NL operation can be expressed as: increased to 16 in all θ, ∅ and g to compensate for the
1 X performance loss caused by down-sampling. The up-sampling
yi = f (xi , xj ) g (xj ) , (7) operation (3 × 3 transposed convolution with stride=2) is used
C(x)
∀j before the addition with the residual connection to recover the
where x and y denote the input and output feature maps, shape of feature maps. For regulation, batch normalization and
respectively, i is the index of a position on feature ReLU activation layers are applied after all convolutions in the
P maps, and j
represents all possible positions on x. C(x) = ∀j f (xi , xj ) NL-block.
is the normalization factor. The function g computes the As shown in Eq. 7, the NL operation calculates the cor-
embedded feature representation of the input feature map at relation f (xi , xj ) at all positions, so that it can directly pass
the position j. Here, we use a 3 × 3 convolution for g. The information as well as extract correlation features between any
function f computes the correlation between index i and j, i.e. two positions (∀j) in one operation. It can also be regarded as a
auto-correlation coefficient matrix. Several forms of f can be global convolution with its own self-correlation matrix as the
selected, e.g., Gaussian, embedded Gaussian and dot product kernel, which can cover the whole map, providing a global
[17], and we choose the embedded Gaussian form: view for feature extraction. The correlation in long-distant
T
∅(xj )
positions can be efficiently captured, which makes NL-block
f (xi , xj ) = eθ(xi ) , (8) more suitable for structural feature extraction. In this way, it
where θ and ∅ represent the embedding spaces, and the only takes a few layers to achieve the best results without
potential of abstract feature representation can be explored introducing too many parameters.
in training to achieve better performance than the original
Gaussian form. For more details, we use 3 × 3 convolution B. DS-RefineNet
for both θ and ∅. Note that the Gaussian multiplication can
To further improve the information flow between layers,
be combined with C(x) to exactly form into the expression
we design a new densely connected convolutional network
of softmax activation.
structure by utilizing dense connectivity [18], called DS-
The structure of NL-block in our model is as depicted in
RefineNet.
Fig. 2. The operation involved in a NL-block is given as:
The structure of the proposed DS-RefineNet is shown in
zi = N L(yi ) + xi , (9) Fig. 3. DS-RefineNet is based on the structure of RefineNet
[7] with the same convolutions. Compared to skip connection
where N L(·) stands for NL-Block function and xi denotes implemented in RefineNet, we introduce direct connections
a residual connection. Considering that the matrix product from any layer to its all subsequent layers. This procedure
process may take up a lot of memory, we add down-sampling can be described as:
operations (3×3 convolution with stride=2) to all θ, ∅ and g to
downsize the feature maps, where the multiplied matrix can xl = Hl ([x0 , x1 , · · · , xl−1 ]), l = 1, 2, 3, (10)

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 4

here to reduce the input feature maps. The implementation of

DS-RefineNet has a significant impact on strengthening feature
propagation and encouraging feature reuse. After these two
DS-RefineNet blocks, the feature maps will pass through a
3 × 3 convolution and a sigmoid activation to output the final
reconstructed CSI.
Fig. 3: The architecture of proposed DS-RefineNet
IV. S IMULATION R ESULTS AND A NALYSIS
In this section, to verify the effectiveness of the proposed
CSI compression feedback algorithm, experimental simula-
where Hl denotes the convolution operation,
tions are developed for indoor and outdoor scenarios, re-
[x0 , x1 , · · · , xl−1 ] represents the concatenation of the
spectively, in an FDD massive MIMO system. Comparative
feature map in layers 0, · · · , l − 1 and xl denotes the output
analysis of our proposed scheme with several other methods
of lth convolution operation.
of CSI feedback compression networks is also performed.
The structure of dense connectivity has a better effect on
improving NN’s flexibility than skip connection in RefineNet.
In RefineNet, skip connection only allows the origin input A. Parameter Setting
to access the final output, while in DS-RefineNet, every two To train DS-NLCsiNet, the end-to-end learning for the
layers have directed connections and each layer can access encoder and decoder is applied. The output to DS-NLCsiNet
the origin input as well as the output at any front node. When is Ĥi , which can be expressed as:
facing the NN’s degradation problem, connections are densely
everywhere in the structure of DS-RefineNet, providing nearly Ĥi = f (Hi ; Θ)
all the selections for the data to choose how to flow the = fde (fen (Hi ; Θen ); Θde ), (11)
indispensable layers and jump across those unnecessary ones.
where Hi , Θ = {Θen ; Θde } and f = fde (fen (·)) denote
While in RefineNet, there are only two ways, i.e., data flows
the input channel matrices, parameter set and autoencoder
through all layers or jumps across them by the skip connection.
network, respectively.
All layers are regarded as a whole that the data cannot deal
We use Adam optimizer with default setting to train our
with any single layer, which greatly reduces the network
frameworks,and we choose the mean squared error (MSE) as
efficiency.
loss function, which is given as:
The flexibility of neural networks in DS-RefineNet brings
many benefits. On one hand, by creating a short path from N
1 X 2
early layers to later layers, it can largely alleviate the problem L(Θ) = kf (si ; Θ) − Hi k2 , (12)
N i=1
of vanishing gradient; on the other hand, this connection
mode in DS-RefineNet makes the transmission of features and where k·k2 is the Euclidean norm, and N is the number of
gradients more efficient, and the network is easier to train. samples in the training data.
Each layer can directly access the gradient from the loss We use the COST2100 channel model [19] to generate the
function and the original input signal, resulting in implicit values of H, considering two different scenarios: the indoor
deep supervision. In the structure of decoder, the received picocellular scenario at the 5.3GHz band and the outdoor rural
compressed signal is first input in the non-local block to get scenario at the 300MHz band. The BS uses N et = 32 antennas
a rough reconstruction of CSI. Then the rough reconstruction and Nc = 1024 subcarriers. We reserve the first Nc = 32
is fed into two DS-RefineNet blocks for more detailed refine- columns of the channel matrix H since only they have non-
ments. zero values. Then H is truncated into the shape of 32 × 32.
In every DS-RefineNet block, there are three 3 × 3 convo- We use 100, 000 samples for training, 30, 000 for validation
lutional layers. The batch normalization layer followed with and 20, 000 for testing, respectively. The batch size, epochs
leaky ReLU activation is implemented before each convolu- and learning rate are set as 200, 1000 and 0.001, respectively.
tional layer. For the lth layer, we concatenate multiple inputs
of Hl (·) as a tensor. The first, second and third convolutional B. Complexity Analysis
layer generates 8, 16 and 2 feature maps, respectively. The
final output of DS-RefineNet is the concatenation of each Compared with other state-of-the-art CSI feedback methods,
layer’s output in the channel axis. the complexity analysis of the proposed DS-NLCsiNet is
depicted in Table I, where the number of parameters and
Due to the concatenation operation, the neural network in
MACCs1 stand for space and time complexity, respectively.
the second DS-RefineNet may become too wide. To further
As shown in Table I, our frameworks do not introduce too
reduce the number of feature maps, a transition layer should
many parameters when compared with CsiNet, while greatly
be implemented between two DS-RefineNet blocks. Different
enhancing the recovery quality of CSI. The increase in MACCs
from [18], we use 1×1 convolution as a substitute for the tran-
mainly comes from convolution layers. When CR is relatively
sition layer. Since the architecture of CSI feedback schemes
is simpler than that of computer vision networks [12], the 1 MACC: multiply-accumulate operations. A multiplication operation and
bottleneck before every 3 × 3 convolution will not be utilized an additive operation count for one MACC operation.

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 5

TABLE I: The number of parameters and MACCS

CR 1/4 1/8 1/16 1/32

CsiNet [7] 2,103,904 1,055,072 530,656 268,448

Params
RecCsiNet [10] 28,331,104 22,300,512 19,481,824 18,121,632
ConvlstmCsiNet [11] 28,326,904 22,296,312 19,477,624 18,117,432
NLCsiNet 2,107,684 1,058,852 534,436 272,228
DS-NLCsiNet 2,108,992 1,060,160 535,744 273,536
MACCs CsiNet [7] 4,366,336 3,842,048 3,579,904 3,448,832
RecCsiNet [10] 153,059,328 128,942,080 117,669,888 112,230,400
ConvlstmCsiNet [11] 121,708,544 97,591,296 86,319,104 80,879,616
NLCsiNet 10,424,448 9,375,872 8,851,548 8,589,440
DS-NLCsiNet 11,497,120 10,448,544 9,924,256 9,662,112

TABLE II: NMSE in dB and cosine similarity ρ

small, the amount of computation of convolutional layers is
more than that of dense layers. The model parameters and Indoor Outdoor
CR Methods
MACCs of DS-NLCsiNet are much lower than those of RecC- NMSE ρ NMSE ρ
siNet and ConvlstimCsiNet, which improve the reconstruction CsiNet -17.36 0.99 -8.75 0.91
accuracy at the cost of huge space and time complexity, 1 NLCsiNet -22.51 0.99 -9.02 0.93
4 DS-CsiNet -20.01 0.99 -10.38 0.93
since the dense layers in LSTM cell substantially increase the DS-NLCsiNet -24.99 0.99 -12.09 0.95
amount of computation. CsiNet -12.70 0.96 -7.61 0.88
1 NLCsiNet -13.60 0.98 -7.69 0.89
8 DS-CsiNet -16.00 0.99 -7.66 0.89
C. Comparative Analysis DS-NLCsiNet -17.00 0.99 -7.96 0.90
To gain insight into the proposed DS-NLCsiNet, the fol- CsiNet -8.65 0.93 -4.51 0.79
1 NLCsiNet -9.76 0.95 -4.80 0.80
lowing CSI feedback methods are implemented in the same 16 DS-CsiNet -9.17 0.94 -4.94 0.82
environment for comparison: DS-NLCsiNet -12.93 0.97 -4.98 0.81
• CsiNet [7]: A well-known CSI sensing and recovery CsiNet -6.24 0.89 -2.81 0.67
1 NLCsiNet -7.58 0.92 -3.18 0.71
mechanism that applies RefineNet. 32 DS-CsiNet -7.38 0.91 -3.22 0.70
• NLCsiNet: Instead of using DS-RefineNet, this scheme DS-NLCsiNet -8.64 0.93 -3.35 0.73
only employs NL-blocks with RefineNet.
• DS-CsiNet: Instead of using NL-blocks, this scheme only
employs DS-RefineNet.
• DS-NLCsiNet: This is the proposed DL-based CSI DS-NLCsiNet outperforms the existing DL-based CSI feed-
feedback scheme that combines NL-blocks and DS- back methods in terms of both NMSE and ρ. Compared with
RefineNet. CsiNet, DS-NLCsiNet also provides significant gains, which
Two metrics can be used to evaluate the performance of mainly benefits from the use of NL-blocks in both of the
different CSI feedback architectures as follows: encoder and decoder, and DS-RefineNet in the decoder. In
addition, the performance comparison of DS-CsiNet, NLC-
• Normalized Mean Square Error (NMSE): It quantifies the
siNet and DS-NLCsiNet demonstrates that NL-blocks and
difference between the original channel matrices and the
DS-RefineNet can indeed enhance the performance of CSI
recovered matrices, which can be defined as:
feedback network, respectively.
2)
(
H−H
b Figure 4 plots original and some reconstructed CSI images
2
N M SE = E 2 . (13) in Pseudo-gray at different compression ratios for different
kHk2 CSI feedback schemes. The CSI images are randomly ex-
• Cosine similarity: It evaluates the similarity between tracted from the test dataset in indoor and outdoor scenarios,
the input and the output matrices by calculating cosine respectively. In the first column from left, we show the
similarity of the channel response at each subcarrier, original images of the CSI matrix in the angular-delay domain
which can be defined as: after performing the 2D-DFT and truncation operation. In
the right three columns, we demonstrate some reconstruction
b̃H h̃
 


 1 X Nec h n n


 samples along with the corresponding pseudo-gray plots of
ρ=E . (14) the strength of recovered matrices, which are reconstructed by
 Nc n=1 hn h̃n 2  using different CSI feedback schemes. Obviously, NLCsiNet

 e b̃ 

2
and DS-NLCsiNet both outperform CsiNet, especially at low
We compare DS-CsiNet, NLCsiNet and DS-NLCsiNet with compression ratios. In addition, NLCsiNet and DS-NLCsiNet
CsiNet. The corresponding NMSE and ρ of each network are can recover the CSI in a more accurate way, and can also
summarized in Table II, where the best results are marked in retain some feature which might be lost in CsiNet feedback
bold font. Simulation results demonstrate that our proposed procedure.

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCOMM.2020.3019653, IEEE
Communications Letters
IEEE COMMUNICATIONS LETTERS, VOL. XX, NO. XX, 2020 6

[8] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition, pp. 770–778, 2016.
[9] T. Wang, C. Wen, S. Jin, and G. Y. Li, “Deep learning-based CSI
feedback approach for time-varying massive MIMO channels,” IEEE
Wireless Communications Letters, vol. 8, pp. 416–419, April 2019.
[10] C. Lu, W. Xu, H. Shen, J. Zhu, and K. Wang, “MIMO channel infor-
mation feedback using deep recurrent network,” IEEE Communications
Letters, vol. 23, no. 1, pp. 188–191, 2018.
[11] X. Li and H. Wu, “Spatio-temporal representation with deep neural
recurrent network in MIMO CSI feedback,” IEEE Wireless Communi-
cations Letters, vol. 9, pp. 653–657, May 2020.
[12] Z. Lu, J. Wang, and J. Song, “Multi-resolution CSI feedback with
deep learning in massive MIMO system,” in Proceedings of the IEEE
International Conference on Communications, June 2020.
[13] J. Guo, C.-K. Wen, S. Jin, and G. Y. Li, “Convolutional neural network-
based multiple-rate compressive sensing for massive MIMO CSI feed-
back: Design, simulation, and analysis,” IEEE Transactions on Wireless
Communications, vol. 19, pp. 2827–2840, Apr. 2020.
[14] J. Guo, X. Yang, C.-K. Wen, S. Jin, and G. Y. Li, “DL-based CSI
feedback and cooperative recovery in massive MIMO,” arXiv preprint
arXiv:2003.03303, 2020.
[15] P. Liang, J. Fan, W. Shen, Z. Qin, and G. Li, “Deep learning and com-
pressive sensing-based CSI feedback in FDD massive MIMO systems,”
IEEE Transactions on Vehicular Technology, pp. 1–1, 2020.
[16] J. Guo, J. Wang, C.-K. Wen, S. Jin, and G. Y. Li, “Compression and
acceleration of neural networks for communications,” IEEE Wireless
Communications, vol. 27, pp. 110–117, Aug. 2020.
[17] X. Wang, R. Girshick, A. Gupta, and K. He, “Non-local neural net-
Fig. 4: The absolute value of original and reconstructed CSI images works,” in Proceedings of the IEEE conference on computer vision and
at different compression ratios (top: in indoor picocellular scenario; pattern recognition, pp. 7794–7803, 2018.
[18] G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely
bottom: in outdoor rural scenario)
connected convolutional networks,” in Proceedings of the IEEE confer-
ence on computer vision and pattern recognition, pp. 4700–4708, 2017.
[19] L. Liu, C. Oestges, J. Poutanen, K. Haneda, P. Vainikainen, F. Quitin,
V. C ONCLUSION F. Tufvesson, and P. De Doncker, “The COST 2100 MIMO channel
model,” IEEE Wireless Communications, vol. 19, no. 6, pp. 92–99, 2012.
In this paper, we have proposed a novel DL-based CSI
feedback scheme by utilizing non-local block and dense con-
nectivity in feature extraction and RefineNet modules, respec-
tively. Experimental results demonstrate that DS-NLCsiNet
outperforms existing methods in terms of recovery accuracy
and reconstruction quality. We believe this architecture has
the potential for practical deployment on real FDD MIMO
systems.

R EFERENCES
[1] C. Li, P. Liu, C. Zou, F. Sun, J. M. Cioffi, and L. Yang, “Spectral-
efficient cellular communications with coexistent one- and two-hop
transmissions,” IEEE Transactions on Vehicular Technology, vol. 65,
pp. 6765–6772, Aug 2016.
[2] M. S. Sim, J. Park, C. Chae, and R. W. Heath, “Compressed channel
feedback for correlated massive MIMO systems,” Journal of Communi-
cations and Networks, vol. 18, pp. 95–104, Feb 2016.
[3] P. Kuo, H. T. Kung, and P. Ting, “Compressive sensing based channel
feedback protocols for spatially-correlated massive antenna arrays,”
in 2012 IEEE Wireless Communications and Networking Conference
(WCNC), pp. 492–497, April 2012.
[4] I. Daubechies, M. Defrise, and C. De Mol, “An iterative thresholding
algorithm for linear inverse problems with a sparsity constraint,” Com-
munications on Pure and Applied Mathematics: A Journal Issued by the
Courant Institute of Mathematical Sciences, vol. 57, no. 11, pp. 1413–
1457, 2004.
[5] D. L. Donoho, A. Maleki, and A. Montanari, “Message-passing algo-
rithms for compressed sensing,” Proceedings of the National Academy
of Sciences, vol. 106, no. 45, pp. 18914–18919, 2009.
[6] T. Wang, C.-K. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin, “Deep
learning for wireless physical layer: Opportunities and challenges,”
China Communications, vol. 14, no. 11, pp. 92–111, 2017.
[7] C.-K. Wen, W.-T. Shih, and S. Jin, “Deep learning for massive MIMO
CSI feedback,” IEEE Wireless Communications Letters, vol. 7, no. 5,
pp. 748–751, 2018.

1089-7798 (c) 2020 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.
Authorized licensed use limited to: Cornell University Library. Downloaded on September 09,2020 at 21:23:47 UTC from IEEE Xplore. Restrictions apply.