Methods and systems for source coding using a neural network

Patent metadata

Field	Value
Publication number	US-12530604-B2
Application number	US-202117459606-A
Country	US
Kind code	B2
Filing date	Aug 27, 2021
Priority date	Aug 27, 2021
Publication date	Jan 20, 2026
Grant date	Jan 20, 2026

Abstract

An apparatus for feature-based communications is provided that includes a probabilistic encoder and a transmitter. The probabilistic encoder is configured to encode source information into a set of probability distributions over a latent space. Each probability distribution represents one or more aspects of a subject of the source information. The transmitter is configured to transmit over a transmission channel, to a receiving electronic device, a set of transmission features representing the subject. Each transmission feature provides information about a respective one of the probability distributions in the latent space. The probabilistic encoder is configured to enforce constraints on distribution parameters of the probability distributions over the latent space based on a condition of the transmission channel. Enforcing constraints on the latent space in this manner enables the apparatus to transmit features that are at least as unreliable as the transmission channel.

First claim

The invention claimed is:

1 . An apparatus for feature-based communications, the apparatus comprising:
a probabilistic encoder for encoding source information into a set of probability distributions over a latent space, each probability distribution representing one or more aspects of a subject of the source information; a transmitter for transmitting over a transmission channel, to a receiving electronic device (ED), a set of transmission features representing the subject, each transmission feature providing information about a respective one of the probability distributions in the latent space, the probabilistic encoder being configured to enforce constraints on distribution parameters of the probability distributions over the latent space based on a condition of the transmission channel, wherein the latent space is a Gaussian distributed latent space and the probabilistic encoder is configured to enforce bounds on the means and variances of the probability distributions over the latent space, and wherein the probabilistic encoder is configured to enforce a minimum variance of the probability distributions over the latent space based on a variance value of a Gaussian distributed model of the transmission channel.

2 . The apparatus of claim 1 , wherein the probabilistic encoder is implemented using an encoder deep neural network (DNN), and the probabilistic encoder uses non-linear activation functions to enforce the constraints on the distribution parameters of the probability distributions over the latent space.

3 . The apparatus of claim 2 , wherein the non-linear activation functions enforce the following constraints on a mean value, μ, and a variance value, σ 2 , of each probability distribution in the latent space:

μ
⁢

ϵ

[

μ
min

,

μ
max

]

,

σ
2

⁢

ϵ

[

σ
W
2

,
∞

]

,

wherein μ min and μ max are lower and upper bounds, respectively, on the mean value and σ W 2 is the variance value of the Gaussian distributed model of the transmission channel.

4 . The apparatus of claim 3 , wherein the encoder DNN is trained using a prior distribution and a Kullback-Leibler (KL) divergence loss term that are functions of the variance value, σ W 2 , of the Gaussian distributed model of the transmission channel between the apparatus and the receiving ED.

5 . The apparatus of claim 1 , wherein:
the probabilistic encoder is a first probabilistic encoder for encoding source information into a set of probability distributions over a first latent space to support a first task; the apparatus further comprises a second probabilistic encoder for encoding source information into a set of probability distributions over a second latent space to support a second task, the second task being different from the first task, and dimensionality of the second latent space being different from dimensionality of the first latent space.

6 . The apparatus of claim 1 , wherein the transmitter is configured to:
transmit a first set of transmission features to the receiving ED, each transmission feature in the first set of transmission features providing information about a respective one of the probability distributions in a first subset of less than all of the probability distributions in the latent space; and subsequent to transmitting the first set of transmission features, transmit a second set of transmission features to the receiving ED, each transmission feature in the second set of transmission features providing information about a respective one of the probability distributions in a second subset of the probability distributions in the latent space, the second subset being different than the first subset.

7 . The apparatus of claim 6 , wherein the second subset is non-overlapping with the first subset.

8 . The apparatus of claim 6 , wherein the transmitter is configured to successively transmit different sets of transmission features providing information about different subsets of the probability distributions in the latent space until either a confirmation message is received from the receiving ED or a predetermined number of transmissions have been made.

9 . The apparatus of claim 1 , wherein the transmitter is configured to transmit the transmission features providing information about respective ones of the probability distributions in the latent space without applying channel coding to the transmission features.

10 . The apparatus of claim 1 , wherein the probabilistic encoder is configured to enforce constraints on distribution parameters of the probability distributions over the latent space such that the transmission features each have an entropy that matches or exceeds an entropy of the transmission channel.

11 . A method for feature-based communications, the method comprising:
encoding source information into a set of probability distributions over a latent space, each probability distribution representing one or more aspects of a subject of the source information; transmitting over a transmission channel, to a receiving electronic device (ED), a set of transmission features representing the subject, each transmission feature providing information about a respective one of the probability distributions in the latent space, wherein constraints are enforced on distribution parameters of the probability distributions over the latent space based on a condition of the transmission channel, wherein the latent space is a Gaussian distributed latent space and bounds are enforced on the means and variances of the probability distributions over the latent space, and wherein a minimum variance of the probability distributions over the latent space is enforced based on a variance value of a Gaussian distributed model of the transmission channel.

12 . The method of claim 11 , wherein the encoding is implemented using an encoder deep neural network (DNN), and non-linear activation functions are used to enforce the constraints on the distribution parameters of the probability distributions over the latent space.

13 . The method of claim 12 , wherein the non-linear activation functions enforce the following constraints on a mean value, μ, and a variance value, σ 2 , of each probability distribution in the latent space:

μ
⁢

ϵ

[

μ
min

,

Assignees

• Cavatassi Adam Christian
• Ge Yiqun
• Aurora Harsh
• Huawei Tech Co Ltd

Inventors

• Cavatassi Adam Christian
• Ge Yiqun
• Aurora Harsh

Classifications

• G06N3/02

  Neural networks · CPC title

• H04B7/02

  Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas (RAKE receivers H04B1/7115) · CPC title

• H04L1/0042

  Encoding specially adapted to other signal generation operation, e.g. in order to reduce transmit distortions, jitter, or to improve signal shape (H04L1/0067 takes precedence) · CPC title

• H04L1/0014

  by adapting the source coding · CPC title

• G06N3/048

  Activation functions · CPC title