Federated AI Model Management with Selective Parameter Updates
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI systems face inefficiencies due to separate development and operation, leading to data silos and resource-intensive maintenance, which can be addressed by integrating local and global models on an AI platform using a Machine Learning Operations (MLOps) platform.
Innovation Solution
A method for managing local and global models on an AI platform through a central server and local servers, utilizing federated learning, where local models share a hidden layer architecture, and encoded reference information is used to cluster and update the global model efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all local model parameters are transmitted to the central server for global model updates, then the global model can be comprehensively updated, but network resources are consumed and communication costs increase
Solution Approach 1:
The patent extracts only the necessary parameter information from local models for global model updates, rather than transmitting all parameters. The central server selectively obtains parameter information from local models based on the encoded reference information, thereby reducing network resource consumption while maintaining update effectiveness.
Solution Approach 2:
The patent implements local quality by allowing different local models to share a common hidden layer architecture while maintaining independence in other layers. This enables the global model to be updated with targeted parameter information from each local model, reducing the amount of data transmission required while preserving model-specific characteristics.
2Adaptability or versatility
If multiple global models are maintained for different tasks, then task-specific performance improves, but system complexity increases
Solution Approach 1:
The patent implements universality by designing a global model with a shared hidden layer architecture that can serve multiple tasks. The common hidden layer acts as a universal feature extractor that benefits all tasks, while task-specific adaptations are achieved through selective parameter updates rather than maintaining separate complete models.
Solution Approach 2:
The patent applies the nested doll principle by embedding task-specific model layers within a universal hidden layer architecture. The shared hidden layer contains general-purpose features that are nested within task-specific adaptations, allowing the system to maintain multiple task capabilities without proportionally increasing overall system complexity.
3Ease of operation
If local models are trained independently without coordination, then local learning flexibility is maintained, but data silos form and overall system efficiency decreases
Solution Approach 1:
The patent implements feedback by establishing a coordination mechanism where local models transmit selected parameter information to the central server, which then updates the global model and feeds it back to local servers. This feedback loop enables local models to benefit from system-wide learning while maintaining their operational flexibility and independence.
Solution Approach 2:
The central server acts as an intermediary that coordinates between independently trained local models. It collects selected parameter information from local models, performs global model updates, and distributes updated models back to local servers, thereby preventing data silos while preserving local learning flexibility.
4Loss of energy
If encoded reference information is used to cluster local models, then parameter transmission is reduced, but additional processing time is required for encoding and clustering
Solution Approach 1:
The patent applies preliminary action by encoding reference information from local models before the global model update process. This encoding is performed in advance to facilitate efficient clustering and selective parameter transmission, reducing the time required during the actual update process despite the additional initial processing step.
Data Source
AI summary
Disclosed is a method for managing a local model and a global model on an AI platform, the method performed by one or more processors of a computing device according to an exemplary embodiment of the present disclosure.the method may include: obtaining one or more encoded reference information, generated by encoding task information or data type information related to local training performed by one or more local servers; obtaining at least a part of parameter information of one or more local models locally trained by the one or more local servers; and updating a global model based on the one or more encoded reference information and at least a part of the parameter information of the one or more local models.