Similarity-Based Decentralized Learning for Accurate Network Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional decentralized machine learning approaches result in generic and inaccurate models due to averaging individual outputs, failing to leverage the specificity of local training data from similar computing environments.
Innovation Solution
Select computing networks with similar infrastructures for decentralized learning, using similarity scores to ensure that the global model is highly accurate when applied to those networks, employing techniques like federated learning to combine local training results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional decentralized machine learning approaches combine all local training results by averaging individual outputs, then data privacy is protected, but the resulting trained models become generic and inaccurate for specific use-cases
Solution Approach 1:
The patent applies local quality by selecting and grouping computing networks based on infrastructure similarity, then training decentralized machine learning models specifically for each similarity group. This ensures that each group receives tailored training data and model parameters suited to its specific infrastructure characteristics, rather than applying a generic averaged model to all networks. The local quality principle resolves the contradiction by making the model adaptive to specific use-cases while maintaining the privacy benefits of decentralized learning.
Solution Approach 2:
The patent segments the population of computing networks into multiple similarity groups based on infrastructure characteristics. Each segment is then trained independently with customized model parameters and selected training data relevant to that specific group. This segmentation approach allows the system to produce specialized models for each infrastructure type rather than a single generic model, thereby improving accuracy for specific use-cases while maintaining data privacy through decentralized training.
2Quantity of substance
If decentralized learning trains using data from diverse computing networks, then more training data is available, but the model becomes less accurate for specific network types
Solution Approach 1:
The patent applies local quality by curating and selecting training data specifically relevant to each similarity group's infrastructure characteristics. Rather than using all available data from diverse networks, the system identifies and uses only the data types and characteristics that are relevant to each specific group. This selective approach maintains data volume benefits while ensuring high accuracy for specific network types by matching training data to the target infrastructure.
Solution Approach 2:
The patent changes the parameter of data relevance by dynamically selecting which training data to use based on the specific similarity group being trained. For each group, the system identifies the most relevant data characteristics and parameters specific to that infrastructure type, rather than using a uniform data set. This parameter change allows the system to maintain sufficient training data volume while optimizing for accuracy in specific use-cases.
3Adaptability or versatility
If a global model is trained to be universally applicable, then it can be deployed across different networks, but it lacks specificity and accuracy for individual network infrastructures
Solution Approach 1:
The patent segments the deployment strategy by creating multiple specialized models for different similarity groups rather than one universal model. Each segmented model is optimized for its specific infrastructure type, achieving high accuracy for that group. The segmentation maintains deployability because each group-based model can still be deployed across multiple networks within that group, balancing specificity with scalability.
Solution Approach 2:
The patent applies universality at the group level rather than the individual network level. Each similarity group serves as a universal target, with a single customized model deployed across all networks within that group. This multi-functionality approach allows one model per group to serve multiple networks with similar infrastructures, maintaining deployability while achieving specificity through the group-based customization strategy.
Data Source
AI summary
A method, computer system, and computer program product are provided for decentralized machine learning. A plurality of computing networks are identified by determining that each computing network of the plurality of computing networks satisfies a predetermined number of criteria. A decentralized learning agent is provided to each computing network, wherein the decentralized learning agent is provided with input parameters for training and is trained using training data associated with a computing network to which the decentralized learning agent is provided. A plurality of learned parameters are obtained from the plurality of computing networks, wherein each learned parameter of the plurality of learned parameters is obtained by training the decentralized learning agent provided to each respective computing network. A global model is generated based on the plurality of learned parameters.