Split Neural Beamforming Feedback to Cut Device Load
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing wireless networks face computational and airtime overhead challenges in determining beamforming feedback matrices due to increasing numbers of antennas, which are exacerbated by complex computations and increased data transmission.
Innovation Solution
Dynamically split the computation of beamforming feedback matrices between access points and devices using neural networks, where the access point determines the layer distribution based on available computational resources and antenna numbers, with devices performing initial calculations and communicating intermediate CSI to the access point for final matrix determination.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the device determines the complete beamforming feedback matrix using neural network, then beamforming accuracy is improved, but computational load on device increases
Solution Approach 1:
The patent divides the neural network into two segments: the first portion is executed by the device to generate intermediate CSI, and the second portion is executed by the access point to generate the final beamforming feedback matrix. This segmentation reduces the computational burden on the device while maintaining beamforming accuracy through collaborative processing.
2Measurement precision
If the device transmits complete CSI data to access point, then beamforming feedback accuracy is improved, but airtime overhead increases
Solution Approach 1:
The patent extracts and processes only the essential features needed for beamforming feedback through the neural network's intermediate CSI representation. By processing data through the first portion of the neural network locally, the system extracts key channel characteristics without transmitting the complete raw CSI dataset, reducing airtime overhead while preserving necessary information for accurate beamforming.
3Measurement precision
If more layers of neural network are implemented, then beamforming feedback accuracy is improved, but computational complexity and airtime overhead increase
Solution Approach 1:
The patent segments the deep neural network into multiple portions distributed between the device and access point. The device executes the first portion with fewer layers, reducing its computational complexity, while the access point executes subsequent portions including the final layers needed for accurate beamforming feedback matrix generation. This distribution allows deeper networks to be used without overwhelming device resources.
Data Source
AI summary
The present disclosure describes a wireless network that dynamically divides the computation of a beamforming feedback matrix between devices and access points. According to an embodiment, an access point includes an antenna, one or more memories, and one or more processors communicatively coupled to the one or more memories. A combination of the one or more processors determines a first number of layers of a neural network to be implemented by a device and instructs the device to implement the first number of layers of the neural network. The combination of the one or more processors also receives, from the device, intermediate CSI produced by the first number of layers of a neural network implemented by the device, applies, to the intermediate CSI, a second number of layers of the neural network to produce a beamforming feedback matrix, and adjusts the antenna based on the beamforming feedback matrix.