Distributed Training Host Aggregating Sub-Models for Meta-Model Optimization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current communication networks lack mechanisms for implementing distributed learning, specifically federated learning, which is essential for optimizing machine learning models across different network elements, leading to suboptimal performance due to separate training of models at each instance without data aggregation.
Innovation Solution
The proposed solution involves a Distributed Training Host (DTH) that trains and generates machine learning sub-models, while a Meta Training Host (MTH) aggregates these sub-models into a meta-model by specifying training requirements, checking maturity conditions, and combining training data to create a unified model, thereby optimizing model performance across the network.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If each network element trains its own machine learning model independently, then each element can operate autonomously, but the overall system performance is suboptimal due to lack of knowledge aggregation
Solution Approach 1:
The patent combines multiple independently trained sub-models from different network elements into a unified meta-model. The meta-training host aggregates sub-models and their training data, synthesizing them into a comprehensive model that leverages knowledge from all participating elements, thereby improving overall system performance while preserving the autonomy of individual elements during the training phase.
2Reliability
If training data is aggregated from multiple distributed hosts, then model performance improves through knowledge aggregation, but data transmission costs and network load increase
Solution Approach 1:
The patent extracts only the essential components needed for model aggregation - specifically the trained sub-models and their associated training data - from distributed hosts and transmits them to the meta-training host. This selective extraction minimizes data transmission requirements while still achieving effective knowledge aggregation, reducing network load compared to transmitting complete datasets or entire model architectures.
3Reliability
If a centralized system aggregates all training data and models, then optimal global performance is achieved, but system complexity and computational requirements increase significantly
Solution Approach 1:
The patent segments the model training and aggregation process into distinct functional components: distributed sub-model training at individual network elements, metadata collection, and meta-model synthesis at a separate meta-training host. This segmentation distributes computational complexity across multiple nodes rather than concentrating it in a single centralized system, making the overall system more manageable while achieving global optimization.
Data Source
AI summary
It is provided a method comprising: monitoring if a request to train a machine learning sub-model is received from a meta-training host; generating training data; training the machine learning sub-model by at least a first subset of the training data if the request is received and at least the first subset of the training data is generated; checking if a predefined condition related to the machine learning sub-model is fulfilled; providing the trained machine learning sub-model and at least a second subset of the training data to the meta-training host if the condition is fulfilled.