Wireless Neural Network Partitioning for Reduced UE Storage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current wireless communication systems face challenges in managing dynamic environments that attenuate or block signals, undermining established wireless channel measuring and reporting mechanisms, necessitating improved techniques for configuring machine learning models to optimize resource use.
Innovation Solution
A base station dynamically configures user equipment with specific neural network functions and machine learning models for performing wireless communications management procedures, allowing flexible and adaptive use of these models without requiring full storage on the UE, and enabling selective model updates.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If machine learning models are stored fully on user equipment for wireless communications management, then model availability and execution speed are improved, but device storage requirements and complexity increase significantly
Solution Approach 1:
The patent segments the machine learning model into two parts: model structure information (architecture, hyperparameters) stored on the UE, and model parameters (weights, biases) stored and managed by the network entity. This segmentation allows the UE to have lightweight storage requirements while still enabling fast local execution by downloading only necessary parameter updates.
Solution Approach 2:
The network entity acts as an intermediary that stores and manages the complete machine learning models, providing model structure information to UEs and enabling parameter updates without requiring UEs to store full models. This intermediary approach balances storage requirements with execution capabilities.
2Adaptability or versatility
If machine learning models are updated frequently to adapt to dynamic wireless environments, then adaptability and performance are improved, but signaling overhead and network resources increase
Solution Approach 1:
The patent extracts only the necessary model parameters (weights and biases) for transmission between the network entity and UE, rather than transmitting complete models. This extraction approach enables frequent updates for environmental adaptability while minimizing signaling overhead by transmitting only the changed parameter subsets.
Solution Approach 2:
The system implements dynamic configuration where the network entity can selectively update model parameters based on current wireless conditions, allowing the model to adapt to changing environments. The configuration includes dynamic model structure information and parameters that can be adjusted without retransmitting the entire model.
3Measurement precision
If comprehensive machine learning models are deployed on user equipment to handle complex wireless scenarios, then processing capability and accuracy are improved, but power consumption and computational burden increase
Solution Approach 1:
The patent implements partial action by having the UE store only model structure information and receive parameter updates on demand, rather than maintaining complete models locally. This approach provides sufficient processing capability for accurate wireless channel measurements while avoiding the excessive power consumption associated with storing and processing large complete models.
Data Source
Figure 1
Figure 2
Figure 3A~3D
AI summary
Certain aspects of the present disclosure provide techniques and apparatus for determining neural network functions (NNFs) and configuring and using corresponding machine learning (ML) models for performing one or more ML-based wireless communications management procedures. An example method performed by a user equipment includes transmitting, to a base station (BS), UE capability information indicating at least one radio capability of the UE and at least one machine learning (ML) capability of the UE and receiving, from the BS based on the UE capability information, ML configuration information indicating at least one neural network function (NNF) and at least one ML model corresponding to the at least one NNF.