Page 45 - ITU Journal Future and evolving technologies Volume 2 (2021), Issue 6 – Wireless communication systems in beyond 5G era

ITU Journal on Future and Evolving Technologies, Volume 2 (2021), Issue 6

DEEP EXTENDED FEEDBACK CODES

2
1
1
1
Anahid Robert Safavi , Alberto G. Perotti , Branislav M. Popović , Mahdi Boloursaz Mashhadi , Deniz Gündüz 2
1 Radio Transmission Technology Laboratory, Huawei Technologies Sweden AB, Kista 164‑94, Sweden,
2 Information Processing and Communications Laboratory, Department of Electrical and Electronic Engineering, Imperial
College London, London SW7 2BT, U.K.
NOTES: Corresponding author: Alberto G. Perotti, alberto.perotti@huawei.com
Anahid Robert Safavi is now with the Wireless Network Algorithm Laboratory, Huawei Technologies Sweden AB.

Abstract – A new Deep Neural Network (DNN)‑based error correction encoder architecture for channels with feedback,
called Deep Extended Feedback (DEF), is presented in this paper. The encoder in the DEF architecture transmits an informa‑
tion message followed by a sequence of parity symbols which are generated based on the message as well as the observations
of the past forward channel outputs sent to the transmitter through a feedback channel. DEF codes generalize Deepcode [1]
in several ways: parity symbols are generated based on forward channel output observations over longer time intervals in
order to provide better error correction capability; and high‑order modulation formats are deployed in the encoder so as to
achieve increased spectral ef iciency. Performance evaluations show that DEF codes have better performance compared to
other DNN‑based codes for channels with feedback.

Keywords – Deep learning, error correction, feedback, ultra‑reliable

1. INTRODUCTION no-
toriously icult problem. Several coding methods for
The ifth generation (5G) wireless cellular networks’ New
channels with feedback have been proposed; see for ex‑
Radio (NR) access technology has been recently speci‑
ample [7,8,9,10,11]. However, all known solutions either
rd
ied by the 3 Generation Partnership Project (3GPP). NR
do not approach the performance predicted in [6] or ex‑
already ills demanding requirements of throughput,
hibit unaffordable complexity. Promising progress has
reliability and latency. However, new use cases stem‑
been made recently by applying Machine Learning (ML)
ming from new application domains (such as industrial
methods [1], where both encoder and decoder are imple‑
automation, vehicular communications or medical appli‑
mented as two separate Deep Neural Networks (DNNs).
cations) call for further signi icant enhancements. For in‑
The DNNs’ icients are determined through a joint
stance, some typical Industrial Internet of Things (IIoT)
encoder‑decoder training procedure whereby encoder
applications would need considerably higher reliability
and decoder in luence each other. In that sense, the cho‑
and shorter transmission delay compared to what 5G/NR sen decoder structure has impact on the resulting code –
can provide nowadays.
a previously unseen feature. Known DNN‑based feedback
codes [1] use different recurrent Neural Network (NN) ar‑
Error correction coding is a key physical layer functional‑
chitectures, Recurrent NNs (RNNs) and Gated Recurrent
ity for guaranteeing the required performance levels. In
Units (GRUs) are used in [1]; Long‑Short Term Memory
conventional systems, error correction is accomplished
(LSTM) architectures have been mentioned in a preprint
by linear binary codes such as polar codes [2], Low Den‑ of [1] as a potential alternative to RNNs for the encoder.
sity Parity Check (LDPC) codes [3] or turbo codes [4], pos‑
sibly combined with retransmission mechanisms such as A new DNN‑based code for channels with feedback called
Hybrid Automatic Request (HARQ) [5]. HARQ performs Deep Extended Feedback (DEF) code is presented in this
an initial transmission followed by a variable number of paper. The encoder transmits an information message
subsequent incremental redundancy transmissions un‑ followed by a sequence of parity symbols which are gen‑
til the receiver noti ies successful decoding to the trans‑ erated based on the message and on observations of the
mitter. Short Acknowledgment (ACK) or Negative ACK past forward channel outputs obtained through the feed‑
(NACK) messages are sent through a feedback channel in back channel. Known DNN‑based codes for channels with
order to inform the transmitter about decoding success. feedback [1] compute their parity symbols based on the
By usage of simple ACK/NACK feedback messages, con‑ information message and on the most recent information
ventional HARQ practically limits the gains that could po‑ received through the feedback channel. The DEF code is
tentially be obtained by an extensive and more ef icient based on feedback extension, which consists of extending
use of the feedback channel. Codes that make full use of the encoder input so as to comprise delayed versions of
feedback potentially achieve improved performance com‑ feedback signals. Thus, the DEF encoder input comprises
pared to conventional codes, as predicted in [6]. the most recent feedback signal and a set of past feed‑
back signals within a given time window. A similar

© International Telecommunication Union, 2021 33