Edge Neural Network Breeding for Drift Prevention
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for continually training neural networks in edge computing environments face challenges such as network strain, bandwidth issues, and unpredictable behavior due to 'Galápagos Syndrome', where edge servers drift and diverge from each other.
Innovation Solution
Deploying multiple copies of a centralized neural network to edge servers, where each copy is independently trained and periodically sent to a cloud-based data center for neural network breeding, updating the centralized network without sending input data back to the cloud, thus reducing network strain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Duration of action of stationary object
If neural networks are independently trained at each edge server, then training continues without interruption, but edge servers drift and diverge from each other causing unpredictable behavior
Solution Approach 1:
The patent implements periodic synchronization where edge neural networks are sent back to the cloud data center at regular intervals for breeding operations. This periodic action maintains continuous training at edge servers while preventing drift through scheduled centralization, resolving the contradiction between uninterrupted training and predictable behavior.
2Reliability
If all input data is sent from edge servers to the cloud for training, then the centralized neural network can be continuously updated, but network bandwidth is consumed and network strain increases
Solution Approach 1:
The patent extracts only the essential training artifacts (trained neural network copies and fitness measures) from edge servers and sends them to the cloud, rather than transmitting all raw input data. This extraction approach enables centralized network updates while dramatically reducing network bandwidth consumption and strain.
Solution Approach 2:
The patent uses copies of trained neural networks and fitness measures as surrogates for the actual training data. Instead of sending original input data, the system transmits replicated network models and their performance metrics, enabling centralized updates with minimal network resource usage.
3Adaptability or versatility
If multiple copies of neural networks are deployed to edge servers for independent training, then training diversity increases, but maintaining consistency across all copies becomes complex
Solution Approach 1:
The patent implements a feedback mechanism where fitness measures from each edge server's neural network performance are sent to the cloud data center. This feedback enables the breeding operation to selectively combine successful training outcomes while maintaining consistency across the centralized network, balancing training diversity with uniformity.
Data Source
AI summary
Continual neural network training in an edge computing environment includes deploying a plurality of copies of a centralized neural network respectively to a corresponding plurality of edge servers, wherein each of the copies of the centralized neural network is independently operated and trained at a respective one of the edge servers based on inputs received at that edge server to create independently trained neural networks. At periodic intervals, copies of the independently trained neural networks and a corresponding fitness measure for each of the copies of the independently trained neural networks are sent from the plurality of edge servers to a cloud-based data center. The centralized neural network is updated at the cloud-based data center, including performing neural network breeding based on the copies of the independently trained neural networks sent from the plurality of edge servers.