Asynchronous Machine Learning Model Training Across Client Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional machine learning systems face challenges in generating accurate models without exposing private client data, lack flexibility in adapting to changes in client-data privacy and sharing settings, and inefficiently consume computing resources due to centralized training approaches.
Innovation Solution
The asynchronous training system trains machine learning models across client devices using local versions, sending global parameters for modification and receiving modified parameter indicators, allowing for adaptive and private training without direct data transmission, thereby distributing the training workload and reducing resource consumption.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional machine learning systems collect and utilize private digital information from client devices to train models on central servers, then model accuracy is improved, but client data privacy is compromised
Solution Approach 1:
The patent extracts only the necessary model parameters from the training process and transmits them to client devices for local training. Instead of centralizing all client data on servers, the system extracts model architecture parameters and sends them distributed to clients, who perform training locally and return only updated parameter values. This extraction approach maintains model accuracy while eliminating the need to collect and store private client data centrally.
Solution Approach 2:
The patent introduces model parameters as an intermediary between the central server and client data. Rather than directly accessing client data, the server communicates through parameter transmissions that enable local training. The parameters act as a mediator that carries training instructions to clients and collects training results without exposing private data, thus resolving the privacy-accuracy contradiction.
2Ease of manufacture
If conventional machine learning systems use centralized training approaches, then model training can be performed, but computing resources on central servers are inefficiently consumed
Solution Approach 1:
The patent segments the centralized training process into distributed local training tasks. Instead of concentrating all training computations on central servers, the system divides the training workload by sending model parameters to multiple client devices, which independently perform training computations on their local data. This segmentation distributes computational energy consumption across many devices, dramatically reducing server resource usage while maintaining training capability.
Solution Approach 2:
The patent enables client devices to perform self-service training by executing model training locally using their own computational resources. Each client device independently trains the model on its local data without requiring server processing power, thereby making the system self-sufficient at the edge and eliminating the need for energy-intensive centralized computation.
3Ease of operation
If conventional machine learning systems use rigid centralized models, then training can be controlled centrally, but the system lacks flexibility to adapt to changes in client-data privacy settings or device availability
Solution Approach 1:
The patent transforms the rigid centralized training model into a dynamic distributed system. Client devices can dynamically join or leave the training process based on their availability, and the system adapts to changing privacy settings by allowing clients to control their own participation. The distributed architecture enables flexible, real-time adaptation without requiring rigid centralized control, resolving the contradiction between operational control and adaptability.
Data Source
AI summary
This disclosure relates to methods, non-transitory computer readable media, and systems that asynchronously train a machine learning model across client devices that implement local versions of the model while preserving client data privacy. To train the model across devices, in some embodiments, the disclosed systems send global parameters for a global machine learning model from a server device to client devices. A subset of the client devices uses local machine learning models corresponding to the global model and client training data to modify the global parameters. Based on those modifications, the subset of client devices sends modified parameter indicators to the server device for the server device to use in adjusting the global parameters. By utilizing the modified parameter indicators (and not client training data), in certain implementations, the disclosed systems accurately train a machine learning model without exposing training data from the client device.