Method for aligning gradient symbols by using bias regarding aircomp in signal amplitude range of receiver

Patent metadata

Field	Value
Publication number	US-12418333-B2
Application number	US-202018024962-A
Country	US
Kind code	B2
Filing date	Sep 7, 2020
Priority date	Sep 7, 2020
Publication date	Sep 16, 2025
Grant date	Sep 16, 2025

Abstract

Disclosed is a method for aligning gradient symbols by using bias regarding AirComp in a signal amplitude range of a receiver. A method for aligning gradient symbols according to an embodiment of the present disclosure comprises the steps of: performing iteration of gradient values included in a first planar vector and applying clipping and bias values so as to obtain a second planar vector having gradient values in which symbols are aligned; and transmitting the second planar vector together with channel information to a server in the form of AirComp. Various embodiments of the present disclosure may be linked to an artificial intelligence module, a drone (unmanned aerial vehicle, UAV), a robot, an augmented reality (AR) device, a virtual reality (VR) device, a device related to a 5G service, and the like.

First claim

The invention claimed is:

1. A method, by a terminal, of performing federated learning in a wireless communication system, the method comprising:
receiving, from a server, a reference signal for channel state information (CSI);
obtaining the CSI based on the reference signal;
obtaining a second flatten vector having gradient values in which signs are aligned, by repeating gradient values included in a first flatten vector and applying clipping and bias values; and
transmitting the second flatten vector together with the CSI to the server in the form of AirComp.

2. The method of claim 1 , wherein the bias value is determined by the number of terminals participating in federated learning and statistical characteristics of the gradients of the first flatten vector.

3. The method of claim 2 , wherein the bias value is determined using a value defined in advance as a table in a transceiver.

4. The method of claim 1 , wherein the CSI is subjected to phase compensation.

5. The method of claim 1 ,
wherein the signs include a positive sign and a negative sign,
wherein values having a negative sign in a first partial vector among the gradient values included in the second flatten vector are clipped, and a negative bias is applied to the partial vector, and
wherein values having a positive sign in a second partial vector among the gradient values included in the second flatten vector are clipped, and a positive bias is applied to the partial vector.

6. The method of claim 1 , wherein a signal transmitted in the transmitting in the form of AirComp has a predetermined transmission power.

7. The method of claim 1 , further comprising:
grouping at least one terminal participating in learning among a plurality of random terminals based on the CSI.

8. The method of claim 7 , wherein a signal transmitted by at least one terminal belonging to the same group has contiguous reception power sensitivity.

9. A method, by a server, of performing federated learning in a wireless communication system, the method comprising:
transmitting, to a plurality of terminals, a reference signal for channel state information (CSI);
receiving, form the plurality of terminals, the CSI based on the reference signal;
receiving gradient signals for federated learning transmitted in an AirComp form from the plurality of terminals, a signal includes the CSI and a second flatten vector;
obtaining an aggregated gradient based on the received gradient signals; and
updating parameters of the global model based on the aggregated gradient.

10. The method of claim 9 , wherein the obtaining of the aggregated gradient includes calculating one aggregated gradient by overlapping the gradient signals for each entry.

11. The method of claim 10 , wherein a bias value is eliminated and not present in the aggregated gradient as a result of overlapping for each entry.

12. A non-transitory computer-readable recording medium having a program for executing the method of claim 1 in a computer system recorded thereon.

Assignees

• Lg Electronics Inc

Inventors

• Jeon Kijun
• Lee Sangrim
• Lee Hojae
• Kim Yeongjun
• Kim Sungjin

Classifications

• H04B7/0634

  Antenna weights or vector/matrix coefficients · CPC title

• H04B7/0626 Primary

  Channel coefficients, e.g. channel state information [CSI] · CPC title

• G06N3/0985

  Hyperparameter optimisation; Meta-learning; Learning-to-learn · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0464

  Convolutional networks [CNN, ConvNet] · CPC title