287

15

Deep Learning Techniques for Decoding Polar Codes

Warren J. Gross, Nghia Doan, Elie Ngomseu Mambou, and Seyyed Ali Hashemi
Department of Electrical and Computer Engineering, McGill University, Montreal, Quebec, Canada

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.

15.1 Motivation and Background

Deep learning (DL) (LeCun et al., 2015) has been widely used in digital communications
through applications such as channel decoding, prediction, equalization, modula-
tion/demodulation, detection, quantization, compression, and spectrum sensing (Ibnkahla,
2000). In (O’Shea and Hoydis, 2017), a new approach in communication systems based on DL
was introduced, where the entire channel transmission was abstracted as an autoencoder: that
is, an end-to-end pipeline designed by joint concatenation of transmitters and receivers in a
single process. It has been established that the block error rate (BLER) of traditional commu-
nication systems can be improved through this scheme. Furthermore, this idea of end-to-end
channel modeling was extended to multiple-input and multiple-output (MIMO) systems.
MIMO suffers from interference between channels in a way that obtaining the optimal signal
is very difficult. It has been shown that such systems can be represented as multiple-input and
-output generative adversarial nets (Goodfellow et al., 2014), which can be optimized using a
customized loss function for the joint model. The challenges and opportunities related to DL
for the wireless physical layer were presented in (Wang et al., 2017).
In (Farsad et al., 2018), a DL-based approach for joint source-channel coding of a text was
proposed. This approach is quite similar to the idea presented in (O’Shea and Hoydis, 2017),
where the source is jointly trained with the channel to reduce transmission distortion. This
Machine Learning for Future Wireless Communications, First Edition. Edited by Fa-Long Luo.
© 2020 John Wiley & Sons Ltd. Published 2020 by John Wiley & Sons Ltd.
288 Machine Learning for Future Wireless Communications

architecture makes use of a long short-term memory (LSTM)-based model, as proposed in

(Graves and Schmidhuber, 2005, Hasim et al., 2014), over a binary erasure channel (BEC) to
ensure a gain in word error rate (WER) when compared to separate source-channel word pro-
cessing. However, the fixed-length word-processing scheme does not allow flexibility at the
incoming source. Similarly, an end-to-end neural network (NN) system was proposed for com-
munication over the air between two software-defined radios (SDRs) in (Dörner et al., 2018).
The concept of an end-to-end NN was also extended for MIMO relays in (Sun and Jing,
2012). In this work, channel training and coherent decoding under estimated channel error
were studied for relay networks, and the setup of a single-antenna as transmitter, R distinct
antenna relays, and a single R antenna receiver is considered. Note that many coherent coop-
erative decoding methods assume perfect and global channel state information (CSI), which
is in reality imperfect, especially in cooperative relay networks. Two types of approaches were
considered: decoding where CSI is assumed noise-free, and matched decoding where the noise
estimation is evaluated. Simulations demonstrate that at least 3R symbol intervals of training are
needed for mismatched decoding versus R + 2 for matched decoding to achieve full diversity.
In addition, adaptive decoding was presented to compensate for the complexity in the matched
decoding. Several works concerning NNs applied to MIMO systems can be found in (Wen et al.,
2018, O’Shea and Hoydis, 2017, Samuel et al., 2017).
It was established that DL decoding for linear block codes is equivalent to deriving the max-
imum energy function of a NN (Bruck and Blaum, 1989). In order to maximize the energy
function of a NN, (Bruck and Blaum, 1989) suggested that decoding of FEC codes can be done
through maximizing polynomials over the N-cube for a (N, K) block code, where N is the
length and K is the dimension of the code. With similar logic, (Tallini and Cull, 1995) predicted
that the NP-complete problem of receiving an error-free message through a channel can be
solved through NN. In (Wu et al., 2002), a neural structure was described as a perceptron with
a higher-order polynomial as a discriminant function. A (2m − 1, 2m − 1 − m) Hamming code
was decoded through only m + 1 assigned weights on each perceptron, m being a positive inte-
ger. This architecture combines two layers: the first is made of a set of parity bits, and the second
is a linear classifier. This proves that high-order codes such as Bose Chaudhuri Hocquenghem
(BCH) codes can be learned successfully through a multilayer perceptron (MLP).
In channel coding, DL-based decoders can provide improvements in error probabilities
over conventional decoders. In (Nachmani et al., 2018), the conventional belief propagation
(BP) decoding algorithm is formalized as a partially connected NN. In addition, by assigning
trainable weights to the BP-based NN, neural BP decoders can achieve the same performance
as the conventional sum-product BP decoding algorithm with a significantly smaller num-
ber of iterations. In (Kim et al., 2018), the sequential decoding of turbo and convolutional
codes (Berrou et al., 1993, Viterbi, 1971) was performed through recurrent neural network
(RNN)-based models. It has been demonstrated that a trained RNN architecture can decode
these codes over additive white Gaussian noise (AWGN) channels with performance near
that of the maximum a posteriori (MAP) decoding given by the Bahl-Cocke-Jelinek-Raviv
(BCJR) algorithm and the maximum likelihood (ML) decoding given by Viterbi algorithm for
sequential codes. This result was confirmed in (Kim et al., 2018) through the same architecture
and extended to codes with larger block lengths. In (Bennatan et al., 2018), a syndrome-based
DL technique was proposed to decode linear block codes. For a BCH code of length 127,
an 11-layer vanilla MLP and a RNN of 4 stacks were considered separately to estimate the
channel noise rather than the transmitted codeword. The syndrome decoding and soft channel
reliabilities were helpful to eliminate the problem of overfitting of the training codeword set.
BCH decoding performance was compared under traditional BP decoding, syndrome-based
vanilla MLP, and syndrome-based stacked RNN. Results showed that syndrome-based stacked
 Deep Learning Techniques for Decoding Polar Codes 289

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.

15.2 Decoding of Polar Codes: An Overview

Polar codes are a recent breakthrough in the field of channel coding, as they were proven
to achieve channel capacity with efficient encoding and decoding algorithms (Arıkan, 2009).
Successive cancellation (SC) and BP decoding algorithms are first introduced to decode polar
codes (Arıkan, 2009). Although SC decoding can provide a low-complexity implementation, its
serial nature prevents the decoder from reaching a high decoding throughput. In addition, the
error-correction performance of SC decoding for short to moderate polar codes does not satisfy
the requirements of the fifth generation of cellular mobile communications (5G). SC list (SCL)
decoding was introduced in (Tal and Vardy, 2015) to improve the performance of SC decoding.
SCL can provide a significant improvement in terms of error probability if the decoder is aided
by a cyclic redundancy check (CRC). With these appealing properties, polar codes have been
selected to be used in the enhanced mobile broadband (eMBB) control channel of 5G, together
with a CRC (3GPP, 2018).
Recently, it has been shown that polar codes can also be decoded using off-the-shelf DL
decoders, which may lead to high decoding throughput thanks to their one-shot-decoding
property (Cammerer et al., 2017). In addition, it was observed that by assigning trainable
weights to the unrolled factor graph of polar codes, neural BP decoders can provide a sig-
nificant error-correction performance gain over conventional BP decoding (Nachmani et al.,
2018). Other approaches such as using DL models for channel noise estimation have also
demonstrated great potential for DL techniques when applied to well-established problems in
channel coding (Bennatan et al., 2018, Liang et al., 2018).
In this section, we first provide some basic knowledge about polar codes and conventional
polar decoders. In the next section, several DL-based decoding algorithms and their variants
for polar codes are discussed, followed by a detailed evaluation concerning error-correction
performance and decoding latency of state-of-the-art DL-aided decoders for a 5G polar code.

15.2.1 Problem Formulation of Polar Codes

A polar code (N, K) of length N with K information bits is constructed by applying
a linear transformation to the message word u = {u0 , u1 , … , uN−1 } as x = uG⊗n where
⊗n
x = {x0 , x1 , …
[ , xN−1
] } is the codeword, G is the n-th Kronecker power of the polarizing
1 0
matrix G = , and n = log2 N. The vector u contains a set  of K information bits and a
1 1
set  of N − K frozen bits. The locations of the frozen bits are known to both the encoder and
c

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

15.2.2 Successive-Cancellation Decoding

The SC decoding algorithm can be illustrated on a polar-code factor graph as shown in
Figure 15.1a, where N = 8 and K = 4. The messages are propagated through the SC processing
elements (SCPEs) as shown in Figure 15.1b, where 𝛽t,s denotes a left-to-right message and 𝛼t,s
denotes a right-to-left message of the t-th bit index at stage s of the factor graph.
In SC decoding, the right-to-left messages 𝛼t,s are soft LLR values, and the left-to-right mes-
sages 𝛽t,s are hard decision bits. The messages in the factor graph are updated as
𝛼t,s = f (𝛼t,s+1 , 𝛼t+2s ,s+1 ) , (15.3)
𝛼t+2s ,s = g(𝛼t,s+1 , 𝛼t+2s ,s+1 , 𝛽t,s ) , (15.4)
𝛽t,s+1 = 𝛽t,s ⊕ 𝛽t+2s ,s , (15.5)
𝛽t+2s ,s+1 = 𝛽t+2s ,s , (15.6)
where
( ( ) ( ))
a b
f (a, b) = 2arctanh tanh tanh
2 2
≈ sgn (a) sgn (b)min (|a|, |b|) , (15.7)
g(a, b, c) = b + (1 − 2c)a, (15.8)
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)

𝛼t ,s , 𝛽t ,s 𝛼t,s + 1, 𝛽t,s + 1

𝛼t + 2s,s, 𝛽t + 2s,s 𝛼t + 2s,s + 1, 𝛽t + 2s,s + 1

(b)

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.

15.2.3 Successive-Cancellation List Decoding

SCL decoding improves the error-correction performance of SC decoding by running multiple
SC decoders in parallel. Instead of using Eq. (15.9) to estimate u as in SC decoding, each bit
is estimated considering both its possible values 0 and 1. Therefore, at each bit estimation, the
number of candidates doubles. In order to constrain the high complexity of SCL decoding, at
each bit estimation, a set of only L candidates are allowed to survive, based on a path metric
that is calculated as (Balatsoukas-Stimming et al., 2015, Hashemi et al., 2016)
∑
t
PMt𝓁 = ln(1 + e−(1−2û j𝓁 )𝛼j,0𝓁 ) , (15.10)
j=0

1∑
t
≈ sgn(𝛼j,0𝓁 )𝛼j,0𝓁 − (1 − 2û j𝓁 )𝛼j,0𝓁 , (15.11)
2 j=0

where 𝓁 is the path index and û j𝓁 is the estimate of bit j at path 𝓁.

15.2.4 Belief Propagation Decoding

Figure 15.2a demonstrates BP decoding on the factor graph representation of (8, 5). The
messages are iteratively propagated through the BP processing elements (BPPEs) (Arıkan,

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

and for the left-to-right messages is

{
i
rt,k i
= f (rt,s , lj,k
i i
+ rj,s ),
(15.13)
i
rj,k i
= f (rt,s , lt,k
i i
) + rj,s ,
where j = t + 2s , k = s + 1.
The BP decoding performs a predetermined Imax update iterations, where the messages are
propagated through all BPPEs in accordance with Eqs. (15.12) and (15.13). Initially, for 0 ≤ t <
N and ∀i ≤ Imax , lt,n
i i
are set to the received channel LLR values LLRx , and rt,0 are set to the LLR
values of the information and frozen bits as
{
0, if ut ∈ ,
LLR∪c = (15.14)
+∞, if ut ∈ c .
All the other left-to-right and right-to-left messages of the BPPEs at the first iteration are set
to 0. After running Imax iterations, the decoder makes a hard decision on the LLR values of the
t-th bit at the information bit stage to obtain the estimated message word as
{
Imax Imax
0, if rt,0 + lt,0 ≥ 0,
û t = (15.15)
1, otherwise.

15.3 DL-Based Decoding for Polar Codes

As mentioned earlier, this section is devoted to the use of DL in decoding polar codes with
emphasis on off-the-shelf DL decoders and DL-aided decoders by addressing their working
principles, algorithm details, and performance evaluations.

15.3.1 Off-the-Shelf DL Decoders for Polar Codes

In (Gruber et al., 2017), it was shown that a MLP decoder can generalize structured codes, e.g.
polar codes, more effectively than random codes, where the MAP performance was obtained
for structured codes but not for random codes. However, the considered code length in (Gruber
et al., 2017) is limited to 16. A more detailed investigation was carried out in (Lyu et al., 2018),
where a comparison was performed between different off-the-shelf network models including
MLP, a convolutional neural network (CNN), and a RNN for polar codes. The RNN model
outperforms the MLP and CNN models in terms of error-correction performance, at the cost
 Deep Learning Techniques for Decoding Polar Codes 293

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.

15.3.2 DL-Aided Decoders for Polar Codes

15.3.2.1 Neural Belief Propagation Decoders
In contrast to off-the-shelf DL decoders, another approach is to exploit domain knowledge to
design DL-aided decoders. In (Nachmani et al., 2018), a deep neural BP decoder is proposed
to improve conventional BP decoding where trainable weights are assigned to the edges of the
unrolled factor graph. The deep network in this case is constrained to be a partially connected
NN, and its inference functions resemble the operations of conventional BP decoding. This idea
was first evaluated on high-density parity check (HDPC) codes in (Nachmani et al., 2018) and
then on polar codes in (Doan et al., 2018b). It was observed that the trainable weights help the
conventional BP decoding to mitigate the detrimental effects of the code’s short cycles, which
are often found in practical linear codes.
Another problem associated with BP decoding is the costly sum-product (SP) algorithm
(Ryan and Lin, 2009). Instead, a low complexity min-sum (MS) algorithm is used in practical
applications (Ryan and Lin, 2009). However, the MS algorithm also introduces decoding
errors due to its poor estimation compared to the SP algorithm. With the objective of tackling
this challenge, neural offset min-sum (NOMS) decoding was proposed in (Lugosch and
Gross, 2017). The NOMS algorithm trains offset parameters and uses them to correct the MS
approximation. It should be noted that NOMS decoding only requires additions, which makes
this decoder particularly attractive for hardware implementation.
294 Machine Learning for Future Wireless Communications

Neural BP Iteration 1 Neural BP Iteration 2

Right-to Left Pass Left-to Right Pass Right-to Left Pass

Stage Stage Stage Stage Stage Stage Stage Stage
2 1 0 0 1 2 1 0

LLRx 1 1 LLRAUAc 2 LLRx 2 2

l2 l1 r1 l2 l1
2
1
l0 l0

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

• NOMS (Lugosch and Gross, 2017)

{i i
lt,s = sgn(lt,k i−1
)sgn(rj,s i
+ lj,k i
) max (0, min (|lt,k |, |rj,s
i−1 i
+ lj,k |) − 𝑤0 ) ,
(15.18)
i
lj,s i
= sgn(lt,k i−1
)sgn(rt,s i
) max (0, min (|lt,k |, |rt,s
i−1
|) − 𝑤3 ) ,
{ i i
rt,k = sgn(rt,s i
)sgn(lj,k i
+ rj,s i
) max (0, min (|rt,s |, |lj,k
i i
+ rj,s |) − 𝑤6 ) ,
(15.19)
i
rj,k i
= sgn(rt,s i
)sgn(lt,k i
) max (0, min (|rt,s |, lt,k
i
|) − 𝑤9 ) + rj,s
i
,
• NNMS (Xu et al., 2017)
{i
lt,s = 𝑤0 f (lt,k
i
, rj,s
i−1 i
+ lj,k ),
(15.20)
i
lj,s = 𝑤3 f (lt,k
i
, rt,s
i−1 i
) + lj,k ,
{ i
rt,k = 𝑤6 f (rt,s
i
, lj,k
i i
+ rj,s ),
(15.21)
i
rj,k = 𝑤9 f (rt,s
i
, lt,k
i i
) + rj,s ,
where 𝑤m ∈ ℝ (0 ≤ m ≤ 11) are the trainable weights.
Optimizing the weights of the neural BP decoders, as depicted in Figure 15.3, can be done
through backpropagation in order to minimize the following objective function
∑ ∑
Imax n−1
HCE (ĥ s , hs ),
i
Loss = (15.22)
i=1 s=0

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

required to obtain the decoded message bits.

15.3.2.2 Joint Decoder and Noise Estimator

In (Liang et al., 2018) a CNN-based noise estimator is used to remove interference noise
between channels. The noise estimator is then coupled with a conventional BP decoder.
Although the proposed iterative denoising-decoding approach in (Liang et al., 2018) showed
a significant performance gain in the case of strong correlations between channels, when
there is no correlation, this approach only achieved a negligible error probability gain over the
conventional BP decoder.
A multiplicative noise model was introduced in (Bennatan et al., 2018) based on the
decoding syndrome. This noise model has two advantages compared to that of the additive
noise-estimation model used in (Liang et al., 2018). First, it reduces the regression problem of
the noise model in (Liang et al., 2018) to a classification problem, i.e. only learn to estimate
the sign of the noise instead of the actual noise value, which is more feasible given the high
dimensional state of the model inputs. Second, the multiplicative noise-estimation model also
preserves the code symmetry conditions (Richardson and Urbanke, 2008), allowing for the use
of all-zero codewords during training.
Figure 15.4 depicts a joint decoder-noise estimator approach derived from (Liang et al., 2018,
Bennatan et al., 2018), where the noise estimator is a DL model such as NN, CNN, or RNN.
Unlike off-the-shelf DL decoders, which try to directly predict the message words given the
channel LLRs, the joint decoder and noise estimator approach only utilizes off-the-shelf DL
models to denoise the channel LLR values, while the main decoding algorithm is still carried
out by conventional BP decoding.
296 Machine Learning for Future Wireless Communications

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

3.5 4 4.5 5 5.5 6 6.5 7 7.5

Eb /N0 [dB]

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.

Table 15.1 Performance gain in dB when compared with BP5 at FER=10−5 .

SCL32 BP30 NOMS NNMS NNMS-RNN BP5-MLP-BP5 BP5-LSTM-BP5

1.5 1.0 0.5 0.7 1.0 0.7 0.9

298 Machine Learning for Future Wireless Communications

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

0 100 200 300 400

Latency [time steps]

Figure 15.6 Latency comparison of various decoders for (64, 32) selected for 5G.

Table 15.2 Number of trainable parameters required for different DL-aided BP

decoders.

NOMS NNMS NNMS-RNN BP5-MLP-BP5 BP5-LSTM-BP5

288 288 864 5403712 3446976

 Deep Learning Techniques for Decoding Polar Codes 299

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.