Decentralized Model Adaptation with Partial Edge Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Computing devices with limited resources face challenges in maintaining the accuracy of inference models due to limited labeled data and the computational expense of updating them, while also needing to protect privacy by minimizing data transfer.
Innovation Solution
A decentralized system where edge devices perform partial calculations for updating inference models using a combination of labeled and unlabeled data, reducing computational load and data exposure.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If labeled data is continuously updated to maintain inference model accuracy, then model accuracy is improved, but computational burden and data storage requirements increase
Solution Approach 1:
The system segments the labeled data into multiple batches and distributes them across multiple edge devices. Each edge device maintains a portion of the labeled data and performs partial calculations, dividing the overall computational burden and data storage requirements across the distributed network while maintaining model accuracy through coordinated updates.
2Measurement precision
If labeled data is transferred frequently to update inference models, then model accuracy is improved, but privacy protection is compromised due to increased data exposure
Solution Approach 1:
The system segments both the data transfer process and the computation process. Only model update parameters and aggregated gradients are transferred between edge devices and central servers, rather than raw labeled data. This segmentation of data and computation protects privacy while maintaining model accuracy through distributed collaborative learning.
Solution Approach 2:
The system introduces cryptographic intermediaries including secret sharing mechanisms and secure multi-party computation protocols. These intermediaries enable collaborative model training across distributed edge devices without requiring direct sharing of raw labeled data, thus protecting privacy while achieving accurate inference models through coordinated updates.
3Measurement precision
If more labeled data is stored at edge devices to improve model accuracy, then inference model performance is improved, but storage requirements and computational resources increase
Solution Approach 1:
The system segments the total labeled dataset and distributes different portions across multiple edge devices. Each device stores only a fraction of the total labeled data, reducing individual storage requirements. The distributed architecture allows the system to leverage the combined data across all devices for model training without requiring any single device to store the complete dataset.
Data Source
AI summary
An edge information handling system (IHS) manager includes a storage for storing a labeled data associated with a use counter; and a vehicle counter; and a processor. The processor is programmed to: update an inference module using the labeled data, determine, after the updating, whether the use counter of the labeled data has exceeded a current use threshold, and in response to the use counter of the labeled data exceeding the current use threshold, initiating replacing of the labeled data with new labeled data from a central IHS. The current use threshold is based on the vehicle counter.