Adaptive Learning Network Using Localized Model Parameter Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing predictive data analysis solutions face inefficiencies and reliability issues in minimizing prediction error, particularly in scaling event prediction systems across multiple devices without compromising data privacy and security.
Innovation Solution
A method involving a global machine learning model initialized across edge computing entities, where local parameters from fine-tuned local models are aggregated to update the global model, allowing continuous distributed learning without sharing local datasets, thus addressing data privacy and security concerns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local datasets are shared across edge computing entities to improve global model accuracy, then predictive accuracy is improved, but data privacy and security are compromised
Solution Approach 1:
The patent extracts only the essential model parameters (weights and biases) from local machine learning models and transmits them to the central server for aggregation, while leaving the actual local datasets retained at each edge computing entity. This extraction approach enables global model improvement without requiring sharing of sensitive raw data, thus resolving the contradiction between predictive accuracy and data privacy/security
Solution Approach 2:
The patent introduces an intermediary aggregation mechanism that combines local model parameters from multiple edge computing entities to create an updated global model. This intermediary process allows knowledge transfer and model improvement without direct data sharing between entities, maintaining data privacy while achieving collaborative learning benefits
2Productivity
If data is transferred between edge computing entities and central server to train models, then model training is improved, but computational and storage efficiency deteriorates
Solution Approach 1:
The patent extracts only the compressed model parameters (weights and biases) from local datasets and transmits them to the central server, rather than transferring entire datasets. This parameter extraction dramatically reduces data transfer volume, computational overhead, and storage requirements while maintaining the ability to train and update models effectively
Solution Approach 2:
The patent transforms the representation of local model knowledge from raw data parameters to condensed model parameters (weights and biases). This parameter transformation enables efficient transmission and aggregation while significantly reducing the computational and storage resources required compared to handling complete datasets
Data Source
AI summary
Various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities, and/or the like for i) initializing a global model with an initial probability distribution that predicts the likelihood of a target classification, and ii) continuously learning from local data sets across a plurality of edge computing entities over time by: a) generating local machine learning models based on the global machine learning model at edge computing entities, b) observing datasets via the local machine learning models, c) aggregating learnings from the local machine learning models, d) updating the global machine learning model to reflect the aggregated learnings, and e) cascading the updated model to propagate the aggregated learnings across the edge computing entities.