Doan Wiley
Doan Wiley
15
In recent decades, the challenge of designing capacity-approaching codes has been one of the
main focuses in digital communications. In this regard, polar codes were introduced as the first
class of error-correcting codes that provably achieve the capacity for any channel at infinite
code length. Recently, deep learning (DL) has shown great potential in a wide range of applica-
tions in digital communications including channel coding for forward error correction (FEC)
codes. Therefore, we believe that a literature review on the intersection between DL and FEC
codes, especially polar codes, can contribute to the coding community. Organized into four
sections, this chapter first provides background and motivation for the use of DL in various
FEC schemes used for wireless communication systems. Section 15.2 introduces polar codes
and their traditional decoding algorithms. In Section 15.3, three major DL-based approaches
for decoding polar codes are presented in terms of decoding performance, algorithm complex-
ity, and decoding latency. The last section of this chapter, Section 15.4, offers further discussion
and conclusions.
RNN performance approached that of the ordered statistics decoding (OSD) of order 2 while
outperforming that of the syndrome-based vanilla MLP at the cost of high complexity and
latency.
In the following section, we will focus on DL techniques used for polar codes as a case study
of DL for FEC codes.
the decoder, and their values are set to 0. The codeword x is then modulated and sent through
the channel where binary phase-shift keying (BPSK) modulation and AWGN channel model
are considered. The soft vector of the transmitted codeword received by the decoder in this
setting can be written as
y = (𝟏 − 2x) + z, (15.1)
290 Machine Learning for Future Wireless Communications
where 𝟏 is an all-one vector of size N, and z ∈ ℝN is the AWGN vector with variance 𝜎 2 and zero
mean. In the log-likelihood ratio (LLR) domain, the LLR vector of the transmitted codeword is
Pr(x = 0|y) 2y
LLRx = ln = . (15.2)
Pr(x = 1|y) 𝜎 2
û 0 x̂ 0
û 1 x̂ 1
û 2 x̂ 2
Bit indices
û 3 x̂ 3
û 4 x̂ 4
û 5 x̂ 5
û 6 x̂ 6
û 7 x̂ 7
(a)
𝛼t ,s , 𝛽t ,s 𝛼t,s + 1, 𝛽t,s + 1
Figure 15.1 (a) Factor graph representation of a polar code with N = 8, K = 4, and {u0 , u1 , u2 } ∈ c ; (b) an SC
processing element (SCPE).
Deep Learning Techniques for Decoding Polar Codes 291
and ⊕ is the bitwise XOR operation. SC decoding is initialized by setting 𝛼t,n = yt , and the
decoding schedule is such that the bits are decoded one by one from u0 to uN−1 . At layer 0, the
elements of u are estimated as
{
0, if ut ∈ c or 𝛼t,0 ≥ 0,
û t = (15.9)
1, otherwise.
1∑
t
≈ sgn(𝛼j,0𝓁 )𝛼j,0𝓁 − (1 − 2û j𝓁 )𝛼j,0𝓁 , (15.11)
2 j=0
Stage indices
s0 s1 s2 s3
û 0 x̂ 0
û 1 x̂ 1
û 2 x̂ 2
Bit indices
û 3 x̂ 3
û 4 x̂ 4
û 5 x̂ 5
û 6 x̂ 6
û 7 x̂ 7
(a)
i−1 i
rt,s rt,s +1 rt,s i rt,s i
lt,s i rt,s +1
lt,s lt,s +1 i lt,s +1
lt,s +1
rj,s+1 i
rj,s i−1
rj,s i rj,s i
lj,s i rj,s +1
lj,s lj,s +1 i−1 lj,s +1
lj,s +1
(b) (c) (d)
Figure 15.2 (a) BP decoding on the factor graph of (8, 5) with {u0 , u1 , u2 } ∈ c ; (b) a BPPE; (c) a right-to-left
message update of a BPPE on an unrolled factor graph; (d) a left-to-right message update of a BPPE on an
unrolled factor graph.
292 Machine Learning for Future Wireless Communications
2010) located in each stage. Each update iteration starts with a right-to-left message pass
that propagates the LLR values from the channel (rightmost) stage to the information bit
(leftmost) stage, and ends with the left-to-right message pass, which occurs in the opposite
order. Figure 15.2b illustrates a BPPE with its corresponding soft messages, where rt,s denotes a
left-to-right message and lt,s denotes a right-to-left message of the t-th bit index at stage s. One
can also apply BP decoding on the unrolled polar-code factor graph (Doan et al., 2018b); thus
the BP iterations in this setup are performed sequentially. Figures 15.2c and 15.2d illustrate
the input and output messages of a BPPE for the right-to-left and left-to-right message updates
on an unrolled factor graph, where the superscript i denotes the iteration number. The update
rule (Arıkan, 2010) for the right-to-left messages of a BPPE is
{
i
lt,s i
= f (lt,k , rj,s
i−1 i
+ lj,k ),
(15.12)
lj,s = f (lt,k , rt,s ) + lj,k ,
i i i−1 i
of the highest decoding complexity. On the other hand, the CNN model provided a performance
gain over the MLP model, with higher computational time. It was also observed in (Lyu et al.,
2018) that the code length impacts the fitting (underfitting versus overfitting) of the deep NN,
and each type of NN has a saturation code length, which is related to the learning capabilities
of the model.
For all the aforementioned off-the-shelf decoders, the networks are formalized to solve a
multi-category classification problem where the correct codewords are used as the training
labels and the corresponding values of LLRx are used as the network’s input. Normally, the
size of the DL decoder scales with the size of the codeword and the natural architecture of
the DL models in use. Finding the network parameters or weights is done by backpropagation
(LeCun et al., 2015) with various optimization methods such as ADAM (Kingma and Ba, 2014)
or RMSPROP (Hinton et al.). Polar decoding under off-the-shelf DL decoders is carried out by
performing the inference phase of the trained DL models, given the channel LLR values.
The main problem associated with off-the-shelf DL decoders when applied to polar codes or
other linear block codes is the curse of dimensionality (Gruber et al., 2017), which states that the
number of required training samples scales exponentially with code length. To overcome this
issue, a scaling approach was introduced in (Cammerer et al., 2017) constraining DL decoders to
only work with sub-codes with small code sizes. Specifically, a partitioned NN (PNN) decoder
for a polar code of size 128 was proposed. The considered polar code was divided into smaller
sub-blocks, and the partitioned DL decoders are trained individually so that the performance
obtained for each sub-block was close to that of MAP decoding. However, the bit-error-rate
(BER) performance of the integrated system is only similar to that of SC decoding. It is worth
mentioning that the latency of the proposed decoder can be reduced as parallel computations
can be exploited for the DL decoders thanks to their one-shot-decoding property.
In (Doan et al., 2018a), a neural decoder was introduced on the basis of the partitioning idea
of (Cammerer et al., 2017). The proposed neural SC (NSC) in (Doan et al., 2018a) preserves the
same decoding performance in terms of BER and frame error rate (FER) as that of PNN, with a
decoding latency improved by 42.5% for a polar code of length 128 and rate 0.5.
0 0 0 l11 l21 r 22 r 21
a a
LLRAUAc a LLRAUAc a
h h h h h h
ĥ 7
1
ĥ 1
1
ĥ 0
1
ĥ 22 ĥ 21 ĥ 20
A polar PE layer
1
x when training,
a x+y x h 1 + ex’ x, y ∈ ℝN
1 − sgn (x)
y when testing.
2 ’
Figure 15.3 A neural BP decoder architecture with two iterations for (8, 5).
In (Nachmani et al., 2018), the architectures from (Nachmani et al., 2018) and (Lugosch and
Gross, 2017) were changed to resemble an RNN by reusing the weights at each iteration; this
reduces their complexity, as fewer parameters are needed. Furthermore, an RNN architecture
based on a successive relaxation technique was constructed, which further improved the pro-
posed RNN-like neural BP decoders.
In (Xu et al., 2017), a neural normalized min-sum (NNMS) decoder was developed, which
adapts the idea in (Nachmani et al., 2018) for the case of polar codes. NNMS also uses a mul-
tiplicative weight to correct the min-sum approximation. This setup can be scaled to large size
polar codes while still maintaining low decoding latency and complexity.
It was shown in (Ren et al., 2015) that the CRC capability is only used as an early stopping cri-
terion with incremental error-correction performance improvement for BP decoding of polar
codes. In (Doan et al., 2018b), by assigning trainable weights to the CRC-Polar concatenated
graph, the proposed decoder has shown a performance gain of 0.5 dB over the conventional
CRC-aided BP at the FER of 10−5 , for a 5G polar code of length 128. The authors in (Doan
et al., 2018b) also derived a general neural BP decoder architecture specified for polar codes.
Figure 15.3 illustrates an example of this architecture where the weights are shared between
each neural BP decoding iteration. The weight-assignment schemes of state-of-the-art neural
BP decoders when applied to the BPPE update functions in Eqs. (15.12) and (15.13) are sum-
marized as follows:
• NNMS-RNN (Nachmani et al., 2018)
{i
lt,s = 𝑤0 f (lt,k
i
, 𝑤1 rj,s
i−1
+ 𝑤2 lj,k
i
),
(15.16)
i
lj,s = 𝑤4 (𝑤3 f (lt,k
i
, rt,s
i−1
)) + 𝑤5 lj,k
i
,
{ i
rt,k = 𝑤6 f (rt,s
i
, 𝑤7 lj,k
i
+ 𝑤8 rj,s
i
),
(15.17)
i
rj,k = 𝑤10 (𝑤9 f (rt,s
i
, lt,k
i
)) + 𝑤11 rj,s
i
,
Deep Learning Techniques for Decoding Polar Codes 295
where HCE is the cross-entropy function, and hs is the correct hard value vector at stage s of
the polar code factor graph that is obtained from the training samples. In the decoding phase,
only the hard estimated values at stage 0 of the polar code factor graph, i.e. ĥ 0 (1 ≤ i ≤ Imax ), is
i
LLR′x BP x̂HT
LLRx
Decoding
Noise q̂
Estimator
|LLR′x|
Abs
Figure 15.4 A joint BP decoder-DL noise estimator as proposed in (Liang et al., 2018). The input of the noise
estimator is the syndrome of the estimated codeword and the magnitude of the estimated channel LLRs
(Bennatan et al., 2018).
The iterative decoding algorithm in Figure 15.4 starts with the first decoding attempt by run-
ning the conventional BP decoding given the channel input LLRx . If the estimated codeword
x̂ and the estimated message word û do not satisfy the G-matrix-based termination condition
(Yuan and Parhi, 2013), the channel LLR values will be denoised and followed by another BP
decoding attempt. Given the syndrome of the conventional BP decoding algorithm, xH ̂ T , where
H is the parity-check matrix of polar codes, and the absolute values of the channel LLR values,
|LLRx |, the DL-based noise estimator predicts the channel noise by estimating its sign values,
̂ The channel LLR values are then updated by flipping the signs at certain positions predicted
q.
by the noise estimator, which results in the denoised channel LLR values, LLR′ x . Another BP
decoding attempt is then carried out given the denoised LLR values. Finally, the decoding is
terminated if the mentioned termination condition is satisfied or a predetermined maximum
number of decoding attempts is reached.
It is worth noticing that the training samples of the noise estimator depicted in Figure 15.4
only include the erroneous syndromes after the first BP decoding, i.e. when x̂ H T is a nonzero
vector, and the corresponding absolute values of the channel LLRs. The label q, q ⊂ {−1, 1}N , is
used as the correct output label, where qj = −1 indicates a flip at the j-th element of the channel
LLRs, while qj = 1 indicates there is no change required for the j-th element.
15.3.3 Evaluation
In this section, we provide a performance comparison in terms of FER for various state-of-
the-art DL-aided BP decoders when applied to a polar code. In addition, the FER performance
of conventional decoders including BP (Arıkan, 2010) and SCL (Balatsoukas-Stimming et al.,
2015, Hashemi et al., 2016) decoding is also plotted. The examined polar code has a code length
of 64, with 32 information bits, and is selected for the eMBB control channel of 5G (3GPP,
2018). We denote BPImax , where Imax ∈ {5, 30}, as the conventional min-sum BP decoder with
Imax iterations, and SCL32 (Tal and Vardy, 2015) as the SCL decoder with a list size of 32. All
the neural BP decoders considered in this section contain five unrolled BP iterations.
We also examine a joint decoder and channel equalizer decoding system by exploiting the idea
proposed in (Liang et al., 2018, Bennatan et al., 2018). We denote the joint decoding systems as
BP5-MLP-BP5 and BP5-LSTM-BP5, where MLP and LSTM refer to the noise-estimator models
using fully connected NNs and stacked LSTM networks, respectively. Note that the network
architectures and parameters of the DL-based noise estimators are adopted from (Bennatan
et al., 2018).
As all the DL-aided decoders considered in this section satisfy the code symmetry conditions
(Richardson and Urbanke, 2008), only all-zero codewords are required for training. The
training data is obtained for various Eb ∕N0 values where Eb ∕N0 ∈ {1, 2, 3, 4, 5, 6, 7, 8}. At each
Eb ∕N0 value, 105 all-zero codewords are obtained using BPSK modulation and the AWGN
Deep Learning Techniques for Decoding Polar Codes 297
10−1
SCL32
BP5
BP30
NOMS [Lugosch and Gross, 2017]
NNMS [Xu et al., 2017]
10−2 NNMS-RNN [Nachmani et al., 2018]
BP5-MLP-BP5 [Liang et al., 2018, Bennatan et al., 2018]
BP5-LSTM-BP5 [Liang et al., 2018, Bennatan et al., 2018]
10−3
FER
10−4
10−5
Figure 15.5 FER performance of various decoders for (64, 32) selected for 5G.
channel model. All neural BP-based decoders are trained for 100 epochs, while BP5-MLP-BP5
and BP5-LSTM-BP5 are trained for 1000 epochs. The mini-batch size is set to 320, and
RMSPROP (Hinton et al.) is used as the optimization algorithm for training. Keras (Chollet
et al., 2015) and TensorFlow (Abadi et al., 2016) are used as our DL frameworks. During
testing, each decoder decodes at least 104 random codewords to obtain at least 50 frames in
errors at each Eb ∕N0 value.
Figure 15.5 illustrates the FER performance of the mentioned decoders. Table 15.1 summa-
rizes the error-correction performance gains of all the decoders in Figure 15.5 with respect to
the baseline BP5 decoder at a target FER of 10−5 . As observed from Table 15.1, NOMS provides
a gain of 0.5 dB compared to the baseline BP5 decoder, while NNMS and BP5-MLP-BP5 both
have a gain of 0.7 dB. On the other hand, NNMS-RNN and BP30 have the same error-correction
performance, which is around 1.0 dB better than that of the baseline BP5, while BP5-LSTM-BP5
is slightly worse than NNMS-RNN. It is worth mentioning that the best neural BP decoder,
NNMS-RNN, is still 0.5 dB away from SCL32.
The decoding latency in terms of time steps for a polar code of size N under BP decoding
with Imax iterations can be calculated as (Arıkan, 2010)
BP Imax = 2Imax log2 (N). (15.23)
As the unrolled factor graphs of the NOMS, NNMS, and NNMS-RNN decoders are equivalent
to that of a traditional BP decoder with five iterations, their decoding latency can also be calcu-
lated by Eq. (15.23). For BP5-MLP-BP5 and BP5-LSTM-BP5 decoders, their decoding latency
in time steps is the sum of the time steps consumed by two successive BP decoders with five
iterations, and a deep NN with a depth of 5. Therefore, their decoding latency can be calculated
as (Liang et al., 2018):
BP Imax -MLP/LSTM-BP Imax = 4Imax log2 (N) + DepthMLP/LSTM . (15.24)
On the other hand, the SCL32 decoder of (Balatsoukas-Stimming et al., 2015) requires (2N +
K − 2) time steps.
Figure 15.6 illustrates the decoding latency in time steps for all the neural decoders considered
in Figure 15.5. It should be noted that by assigning trainable weights to the factor graph of polar
codes, NNMS-RNN with 5 iterations is able to achieve the same error-correction performance
of BP30, which also results in a saving of 300 time steps. In addition, the decoding latency of
BP5-MLP-BP5 and BP5-LSTM-BP5 is 65 time steps greater than that of NNMS-RNN.
Table 15.2 gives a detailed comparison in the number of weights required by differ-
ent DL-aided BP decoders in Figure 15.5. Although it is demonstrated in Figure 15.5 that
off-the-shelf deep networks are able to estimate channel noise, this approach shows inefficiency
since relatively large DL models are required for the task. On the contrary, by incorporating the
conventional BP decoding algorithm to define a constrained network model, NNMS-RNN can
provide a reasonable error probability while only requiring a small number of weights compared
to those of BP5-MLP-BP5 and BP5-LSTM-BP5. Furthermore, although NOMS and NNMS only
BP5-LSTM-BP5 125
BP5-MLP-BP5 125
NNMS-RNN 60
NNMS 60
NOMS 60
BP5 60
BP30 360
SCL32 158
Figure 15.6 Latency comparison of various decoders for (64, 32) selected for 5G.
require 33.33% of the number of weights consumed by NNMS-RNN, the smaller number of
weights results in a considerable error-correction performance loss, as observed in Figure 15.5.
15.4 Conclusions
In this chapter, we have discussed a wide range of fruitful applications of digital communi-
cations where deep learning can play a vital role. We provided an overview of DL techniques
with a focus on FEC codes and examined state-of-the-art DL-aided decoders for polar codes as
our case study. It was demonstrated that off-the-shelf DL decoders can reach MAP decoding
performance for short code lengths and that they enable parallel execution thanks to their
one-shot-decoding property. However, for longer code lengths, off-the-shelf DL decoders
require a training dataset that scales exponentially with the code length. This issue becomes the
main challenge for those decoders in practical applications. On the other hand, domain knowl-
edge can be exploited to design DL-aided decoders, as demonstrated by various neural BP
decoders. It was shown that neural BP decoders can obtain significant decoding performance
gains over their conventional counterpart, while maintaining the same decoding latency.
Future research on applying DL techniques to FEC can be carried out in various directions,
such as designing jointly trained systems of customized DL-aided decoders and neural
channel-noise estimators for various nonlinear communication channels. In addition, the
sequential decoding of linear block codes such as polar codes can be suitably formalized
as a RL problem, thus greatly enabling the applications of state-of-the-art RL algorithms to
FEC. Other approaches may include the use of DL techniques as optimization methods for
well-defined problems of conventional decoding algorithms, whose solutions are obtained
based on an approximation or a massive Monte Carlo simulation.
Bibliography
3GPP. Multiplexing and channel coding (Release 10) 3GPP TS 21.101 v10.4.0. Oct. 2018. URL
http://www.3gpp.org/ftp/Specs/2018-09/Rel-10/21_series/21101-a40.zip.
M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, et al. TensorFlow: A system for large-scale
machine learning. Proceedings of the 12th USENIX Conference on Operating Systems Design and
Implementation, pages 265–283, 2016.
E. Arıkan. Channel polarization: A method for constructing capacity-achieving codes for
symmetric binary-input memoryless channels. IEEE Transactions on Information Theory, 55 (7):
3051–3073, July 2009. ISSN 0018-9448. doi: 10.1109/TIT.2009.2021379.
E. Arıkan. Polar codes: A pipelined implementation. Proceedings of the 4th International
Symposium on Broadband Communications, pages 11–14, 2010.
A. Balatsoukas-Stimming, M.B. Parizi, and A. Burg. LLR-based successive cancellation list
decoding of polar codes. IEEE Transactions on Signal Processing, 63 (19): 5165–5179, Oct 2015.
ISSN 1053-587X. doi: 10.1109/TSP.2015.2439211.
A. Bennatan, Y. Choukroun, and P. Kisilev. Deep learning for decoding of linear codes - a
syndrome-based approach. 2018 IEEE International Symposium on Information Theory (ISIT),
pages 1595–1599, June 2018. ISSN 2157-8117. doi: 10.1109/ISIT.2018.8437530.
C. Berrou, A. Glavieux, and P. Thitimajshima. Near Shannon limit error-correcting coding and
decoding: Turbo-codes. Proceedings of ICC ’93 - IEEE International Conference on
Communications, 2: 1064–1070 vol.2, May 1993. doi: 10.1109/ICC.1993.397441.
300 Machine Learning for Future Wireless Communications
J. Bruck and M. Blaum. Neural networks, error-correcting codes, and polynomials over the binary
n-cube. IEEE Transactions on Information Theory, 35 (5): 976–987, Sept 1989. ISSN 0018-9448.
doi: 10.1109/18.42215.
S. Cammerer, T. Gruber, J. Hoydis, and S.T. Brink. Scaling deep learning-based decoding of polar
codes via partitioning. IEEE Global Communication Conference, pages 1–6, December 2017. doi:
10.1109/GLOCOM.2017.8254811.
F. Chollet et al. Keras. https://keras.io, 2015.
N. Doan, S.A. Hashemi, and W.J. Gross. Neural successive cancellation decoding of polar codes.
2018 IEEE 19th International Workshop on Signal Processing Advances in Wireless
Communications (SPAWC), pages 1–5, June 2018a. ISSN 1948-3252. doi:
10.1109/SPAWC.2018.8445986.
N. Doan, S.A. Hashemi, E.N. Mambou, T. Tonnellier, and W.J. Gross. Neural belief propagation
decoding of CRC-Polar concatenated codes. arXiv preprint arXiv:1811.00124, 2018b.
S. Dörner, S. Cammerer, J. Hoydis, and S.T. Brink. Deep learning based communication over the
air. IEEE Journal of Selected Topics in Signal Processing, 12 (1): 132–143, Feb 2018. ISSN
1932-4553. doi: 10.1109/JSTSP.2017.2784180.
N. Farsad, M. Rao, and A. Goldsmith. Deep learning for joint source-channel coding of text. 2018
IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages
2326–2330, April 2018. ISSN 2379-190X. doi: 10.1109/ICASSP.2018.8461983.
I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y.
Bengio. Generative adversarial nets. Advances in Neural Information Processing Systems 27,
pages 2672–2680, 2014.
A. Graves and J. Schmidhuber. Framewise phoneme classification with bidirectional lstm and other
neural network architectures. Neural Networks, 18 (5): 602–610, 2005. ISSN 0893-6080. doi:
https://doi.org/10.1016/j.neunet.2005.06.042. IJCNN 2005.
T. Gruber, S. Cammerer, J. Hoydis, and S.T. Brink. On deep learning-based channel decoding. 2017
51st Annual Conference on Information Sciences and Systems (CISS), pages 1–6, March 2017.
doi: 10.1109/CISS.2017.7926071.
S.A. Hashemi, C. Condo, and W.J. Gross. A fast polar code list decoder architecture based on
sphere decoding. IEEE Transactions on Circuits and Systems I: Regular Papers, 63 (12):
2368–2380, Dec 2016. ISSN 1549-8328. doi: 10.1109/TCSI.2016.2619324.
S. Hasim, W.S. Andrew, and F. Beaufays. Long short-term memory recurrent neural network
architectures for large scale acoustic modeling. Fifteenth annual conference of the international
speech communication association (INTERSPEECH), 2014.
G. Hinton, N. Srivastava, and K. Swersky. Neural networks for machine learning lecture 6a
overview of mini-batch gradient descent. URL https://cs.toronto.edu/~tijmen/csc321/slides/
lecture_slides_lec6.pdf.
M. Ibnkahla. Applications of neural networks to digital communications–a survey. Signal
processing, 80 (7): 1185–1215, 2000. doi: 10.1016/s0165-1684(00)00030-x.
H. Kim, Y. Jiang, R. Rana, S. Kannan, S. Oh, and P. Viswanath. Communication algorithms via deep
learning. International Conference on Learning Representations (ICLR), 2018.
D.P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521 (7553): 436, May 2015.
F. Liang, C. Shen, and F. Wu. An iterative BP-CNN architecture for channel decoding. IEEE Journal
of Selected Topics in Signal Processing, 12 (1): 144–159, Feb 2018. ISSN 1932-4553. doi:
10.1109/JSTSP.2018.2794062.
Deep Learning Techniques for Decoding Polar Codes 301
L. Lugosch and W.J. Gross. Neural offset min-sum decoding. 2017 IEEE International Symposium
on Information Theory (ISIT), pages 1361–1365, June 2017. ISSN 2157-8117. doi:
10.1109/ISIT.2017.8006751.
W. Lyu, Z. Zhang, C. Jiao, K. Qin, and H. Zhang. Performance evaluation of channel decoding with
deep neural networks. 2018 IEEE International Conference on Communications (ICC), pages
1–6, May 2018. ISSN 1938-1883. doi: 10.1109/ICC.2018.8422289.
E. Nachmani, E. Marciano, L. Lugosch, W.J. Gross, D. Burshtein, and Y. Be’ery. Deep learning
methods for improved decoding of linear codes. IEEE Journal of Selected Topics in Signal
Processing, 12 (1): 119–131, February 2018. ISSN 1932-4553. doi: 10.1109/JSTSP.2017.2788405.
T.J. O’Shea and J. Hoydis. An introduction to deep learning for the physical layer. IEEE
Transactions on Cognitive Communications and Networking, 3 (4): 563–575, Dec 2017. ISSN
2332-7731. doi: 10.1109/TCCN.2017.2758370.
Y. Ren, C. Zhang, X. Liu, and X. You. Efficient early termination schemes for belief-propagation
decoding of polar codes. IEEE 11th International Conference on ASIC, pages 1–4, Nov 2015. doi:
10.1109/ASICON.2015.7517046.
T. Richardson and R. Urbanke. Modern coding theory. Cambridge university press, 2008.
W. Ryan and S. Lin. Channel codes: classical and modern. Cambridge University Press, 2009.
N. Samuel, T. Diskin, and A. Wiesel. Deep MIMO detection. 2017 IEEE 18th International
Workshop on Signal Processing Advances in Wireless Communications (SPAWC), pages 1–5, July
2017. ISSN 1948-3252. doi: 10.1109/SPAWC.2017.8227772.
S. Sun and Y. Jing. Training and decodings for cooperative network with multiple relays and receive
antennas. IEEE Transactions on Communications, 60 (6): 1534–1544, June 2012. ISSN
0090-6778. doi: 10.1109/TCOMM.2012.050912.110380.
I. Tal and A. Vardy. List decoding of polar codes. IEEE Transactions on Information Theory, 61 (5):
2213–2226, May 2015. ISSN 0018-9448. doi: 10.1109/TIT.2015.2410251.
L.G. Tallini and P. Cull. Neural nets for decoding error-correcting codes. IEEE Technical
Applications Conference and Workshops. Northcon/95. Conference Record, pages 89–, Oct 1995.
doi: 10.1109/NORTHC.1995.485019.
A. Viterbi. Convolutional codes and their performance in communication systems. IEEE
Transactions on Communication Technology, 19 (5): 751–772, October 1971. ISSN 0018-9332.
doi: 10.1109/TCOM.1971.1090700.
T. Wang, C. Wen, H. Wang, F. Gao, T. Jiang, and S. Jin. Deep learning for wireless physical layer:
Opportunities and challenges. China Communications, 14 (11): 92–111, Nov 2017. ISSN
1673-5447. doi: 10.1109/CC.2017.8233654.
C. Wen, W. Shih, and S. Jin. Deep learning for massive MIMO CSI feedback. IEEE Wireless
Communications Letters, 7 (5): 748–751, Oct 2018. ISSN 2162-2337. doi:
10.1109/LWC.2018.2818160.
J.L. Wu, Y.H. Tseng, and Y.M. Huang. Neural network decoders for linear block codes.
International Journal of Computational Engineering Science, 3 (3): 235–256, 2002. doi:
10.1142/S1465876302000629.
W. Xu, Z. Wu, Y. Ueng, X. You, and C. Zhang. Improved polar decoder based on deep learning.
2017 IEEE International Workshop on Signal Processing Systems (SiPS), pages 1–6, Oct 2017.
ISSN 2374-7390. doi: 10.1109/SiPS.2017.8109997.
B. Yuan and K.K. Parhi. Architecture optimizations for bp polar decoders. 2013 IEEE International
Conference on Acoustics, Speech and Signal Processing, pages 2654–2658, May 2013. ISSN
1520-6149. doi: 10.1109/ICASSP.2013.6638137.