Neural networks for decoding

Patent metadata

Field	Value
Publication number	US-11568214-B2
Application number	US-201816636128-A
Country	US
Kind code	B2
Filing date	Aug 22, 2018
Priority date	Aug 23, 2017
Publication date	Jan 31, 2023
Grant date	Jan 31, 2023

Abstract

Methods and apparatus for training a Neural Network to recover a codeword of a Forward Error Correction (FEC) code are provided. Trainable parameters of the Neural Network are optimised to minimise a loss function. The loss function is calculated by representing an estimated value of the message bit output from the Neural Network as a probability of the value of the bit in a predetermined real number domain and multiplying the representation of the estimated value of the message bit by a representation of a target value of the message bit. Training a neural network may be implemented via a loss function.

First claim

The invention claimed is:

1. The method for training a Neural Network, NN, to recover a codeword of a Forward Error Correction, FEC, code from a received signal, wherein layers of the NN implement sequential iterations of the Sum Product Algorithm, SPA, and wherein the received signal comprises a transmitted codeword and channel impairments, the method comprising:
inputting to an input layer of the NN a representation of message bits of a transmitted codeword obtained from a received signal;
propagating the representation through the NN;
calculating a loss function; and
optimising trainable parameters of the NN to minimise the loss function;
wherein calculating a loss function comprises, for bits in the transmitted codeword:
representing an estimated value of the message bit output from the NN as a probability of the value of the bit in a predetermined real number domain; and
multiplying the representation of the estimated value of the message bit by a representation of a target value of the message bit.

2. The method as claimed in claim 1 , wherein calculating a loss function further comprises:
averaging over all bits in the transmitted codeword, the values obtained from multiplying, for bits in the transmitted codeword, the representation of the estimated value of the message bit by a representation of a target value of the message bit.

3. The method as claimed in claim 1 , wherein representing an estimated value of the message bit output from the NN as a probability of the value of the bit in a real number domain comprises:
obtaining a probability of the value of the bit from a layer of the NN; and
transforming the obtained probability to a value within the predetermined real number domain.

4. The method as claimed in claim 3 , wherein the predetermined real number domain is [−1, 1] and wherein transforming the obtained probability to a value within the predetermined real number domain comprises performing a linear transformation on the obtained probability.

5. The method as claimed in claim 1 , wherein the representation of the target value of the message bit comprises a value of the message bit after modulation using a modulation technique applied to the transmitted codeword.

6. The method as claimed in claim 1 , wherein calculating a loss function comprises:
calculating the loss function on the basis of an estimated value of the message bit output from an output layer of the NN.

7. The method as claimed in claim 1 , wherein the loss function comprises:

L
f
E

(

p
,
y

)

=

-
1

N

⁢

Σ

n
=
1

N

(

(

1
-

2
⁢

p
⁡
(
n
)

)

⁢

(

-
1

)

y
⁡
(
n
)

)

wherein:
N is the number of bits in the transmitted codeword;
p(n) is the probability of the value of the n th bit of the transmitted codeword output by the NN being 1; and
y(n) is the target value of the n th bit of the transmitted codeword.

8. The method as claimed in claim 1 , wherein calculating a loss function comprises:
calculating the loss function on the basis of estimated values of the message bit output from even layers of the NN.

9. The method as claimed in claim 8 , wherein the loss function comprises:

L
m
E

(

p
,
y

)

=

-
1

MN

⁢

∑

l
=
2

,
4
,
…

2
⁢
M

(

∑

n
=
1

N

(

(

1
-

2
⁢

p
⁡
(

l
,
n

)

)

⁢

(

-
1

)

y
⁡
(
n
)

Assignees

• Ericsson Telefon Ab L M

Inventors

• Tullberg Hugo
• Agrawal Navneet

Classifications

• H03M13/3977

  using sequential decoding, e.g. the Fano or stack algorithms · CPC title

• G06N3/0472 Primary

  Physics · mapped topic

• G06N3/08

  Learning methods · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title