Wireless Federated Learning Bias Removal for Low-Delay Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The integration of artificial intelligence in wireless communication systems faces challenges in implementing federated learning with reduced delay and effective data communication between base stations and terminals, particularly due to biases in gradient sign determination and ambiguity in real number value ranges.
Innovation Solution
A method and apparatus for federated learning in wireless communication systems that involve a base station receiving local parameters from multiple terminals, integrating them by removing biases based on gradient signs, and transmitting the integrated parameter, utilizing a processor to determine and remove biases using removal parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If federated learning is performed by collecting local parameters from multiple terminals and integrating them at a base station, then distributed learning capability and data privacy protection are improved, but communication delay and system complexity increase
Solution Approach 1:
The patent segments the federated learning process into distinct phases: local parameter generation at terminals, bias determination and removal at base station, and integrated parameter transmission. By segmenting the parameter integration process and performing bias removal locally at the base station before aggregation, the system reduces the time required for parameter exchange and accelerates the overall learning convergence.
Solution Approach 2:
The patent applies preliminary action by determining and removing biases from local parameters before the parameter aggregation process. The base station determines removal parameters based on gradient signs in advance, and terminals remove these biases from their local parameters before transmission. This preliminary bias removal prevents accumulation of biased gradients and reduces the number of communication rounds needed for convergence.
2Measurement precision
If bias removal is performed using gradient signs and removal parameters, then learning accuracy and model convergence are improved, but computational complexity and processing time increase
Solution Approach 1:
The patent applies local quality by performing bias removal operations at specific locations in the distributed system - the base station determines removal parameters based on aggregated gradient information, while individual terminals apply the bias removal to their local parameters. This distributed approach with localized processing reduces the computational burden on any single device while maintaining overall learning accuracy.
Solution Approach 2:
The patent changes parameters by using gradient signs and removal parameters to adjust local parameters before aggregation. The base station determines removal parameters based on the signs of received gradients, and terminals use these parameters to modify their local gradients. This parameter transformation approach enables bias removal while maintaining the integrity and convergence properties of the federated learning process.
3Adaptability or versatility
If multiple terminals transmit local parameters to the base station for integration, then model diversity and learning robustness are improved, but reception signal distribution widens and power gain decreases
Solution Approach 1:
The patent applies inversion by reversing the conventional approach: instead of having terminals transmit raw local parameters and having the base station handle all bias removal, the base station first determines removal parameters from aggregated gradient information, then terminals apply these removal parameters to their local parameters before transmission. This inverted sequence narrows the reception signal distribution and improves power gain while preserving model diversity.
Solution Approach 2:
The patent implements feedback by using the base station's determination of removal parameters based on aggregated gradient signs, and feeding this information back to terminals for local parameter adjustment. This feedback mechanism enables terminals to pre-process their parameters with information about the overall gradient distribution, resulting in narrower reception signal distribution and improved reception reliability at the base station.
Data Source
AI summary
Disclosed is a method by which a base station performs federated learning in a wireless communication system. A method, according to one embodiment of the present disclosure, comprises: receiving, by a base station, a plurality of local parameters from a plurality of terminals; obtaining an integrated parameter on the basis of the plurality of local parameters; and transmitting the integrated parameter to each of the plurality of terminals, wherein at least one bias included in the plurality of local parameters is removed by using the plurality of local parameters. The base station/terminals of the present disclosure may be linked to an artificial intelligence, module, an unmanned aerial vehicle (UAV), a robot, an augmented reality (AR) device, a virtual reality (VR) device, a device related to a 6G service, and the like.