Wireless Model Ensembling for Cross-Cell Knowledge Transfer
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional federated learning in wireless communication networks is limited to using terminal data from a single cell, preventing cross-cell knowledge transfer and affecting the performance of neural network models when terminal devices are handed over between cells.
Innovation Solution
Implementing inter-cell knowledge transfer through model ensembling between terminal devices and network devices across different cells, allowing for joint training and updating of neural network models, even when devices are handed over between cells, using methods like conventional federated learning, split learning, or knowledge distillation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional federated learning is used with terminal data from a single cell, then the training process is simple and manageable, but cross-cell knowledge transfer cannot be implemented and model performance deteriorates during handovers
Solution Approach 1:
The patent merges model training across multiple cells by enabling terminal devices to perform model ensembling on models from different network devices (source and target cells). This combining approach allows knowledge transfer between cells, maintaining model performance during handovers while managing complexity through coordinated model exchange between network devices.
2Reliability
If model ensembling is performed on terminal devices across different cells, then inter-cell knowledge transfer is enabled and model performance improves, but communication overhead and system complexity increase
Solution Approach 1:
The patent extracts and transfers only the essential model parameters (weights and biases) between network devices and terminal devices, rather than transferring complete models or raw data. This extraction approach enables knowledge transfer while minimizing communication overhead, as only critical model information needs to be exchanged during handovers.
Solution Approach 2:
The patent uses model copying where terminal devices receive model parameters from network devices, create local copies, and perform ensembling operations. This copying mechanism allows multiple terminal devices to independently perform model ensembling without requiring continuous communication, reducing overall system communication overhead.
3Measurement precision
If terminal devices frequently download updated models during handovers, then model accuracy is maintained, but radio resources are consumed and system efficiency decreases
Solution Approach 1:
The patent performs preliminary model preparation by having network devices send model parameters to terminal devices before handover occurs. Terminal devices then perform model ensembling locally using these pre-received parameters, avoiding the need for frequent model downloads during actual handovers. This preliminary action maintains model accuracy while significantly reducing radio resource consumption.
Solution Approach 2:
The patent enables terminal devices to perform self-service model updating through local model ensembling operations. Instead of requiring network devices to continuously push updated models to terminal devices, the terminal devices autonomously combine received model parameters with their local models, reducing the frequency and volume of radio communications needed for model maintenance.
Data Source
AI summary
Embodiments of this application provide a wireless communication method and apparatus. The method includes the following: a terminal device obtains a first model of a first network device. The terminal device receives first model indication information from a second network device. The first model indication information indicates a second model of the second network device. The terminal device sends second model indication information to the second network device. The second model indication information indicates an ensemble model of the first model and the second model. According to the methods provided in this application, the terminal device performs model ensembling, to implement knowledge transfer between network devices, thereby improving performance of a neural network model.