Distributed AI Training Data Replication via Model Copying
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Distributing large training data sets to remote locations is challenging due to storage requirements and lengthy data transmission times, especially over slow network connections.
Innovation Solution
A computing system that uses machine learning models to generate and replicate distributed training data, where a data generation machine learning model creates multiple versions of training data with corresponding sequence numbers, and a distributed machine learning model trains on these versions to generate identical copies at remote locations without full data transmission.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If training data is transmitted to remote locations, then data availability at multiple locations is improved, but network resource usage and transmission time increase
Solution Approach 1:
Instead of transmitting the entire training data set to remote locations, the patent creates a copy of the machine learning model that can generate training data locally. The model includes a data generation component that replicates training data without requiring full data transmission, thus reducing network resource usage while maintaining data availability.
Solution Approach 2:
The machine learning model acts as an intermediary between the central training data source and remote computing devices. Rather than directly transmitting raw training data, the model receives training data, processes it through its data generation component, and produces replicated training data locally, thereby reducing network bandwidth requirements.
2Reliability
If training data is transmitted to remote locations, then data availability at multiple locations is improved, but transmission time increases
Solution Approach 1:
The machine learning model performs preliminary processing of training data by generating a condensed representation or model parameters that capture the essential characteristics of the training data. This preliminary action allows remote devices to generate training data locally without waiting for full data transmission, significantly reducing transmission time while maintaining data availability.
Solution Approach 2:
The patent creates a computational copy of the training data generation capability rather than copying the entire training data set. This allows remote locations to generate training data on-demand using the transmitted model parameters, eliminating the need for lengthy full data transmissions while ensuring data availability.
3Reliability
If training data is transmitted to remote locations, then data availability at multiple locations is improved, but storage requirements increase
Solution Approach 1:
Instead of storing and transmitting entire training data sets to multiple locations, the patent transmits a compressed copy of the machine learning model that can generate training data. This reduces storage requirements at remote locations while maintaining the ability to access and use training data through local generation.
Solution Approach 2:
The patent extracts the essential information from the training data and encodes it in the machine learning model parameters. This extraction allows the core characteristics of the training data to be transmitted and stored efficiently, while the full training data can be regenerated locally when needed, reducing overall storage requirements.
Data Source
AI summary
Aspects of the disclosure relate to generating and replicating training data. A computing system may generate different versions of training data and corresponding different sequence numbers. A distributed machine learning model may be trained to generate different versions of the training data based on the corresponding different sequence numbers. The training data and identical copies of the distributed machine learning model may be sent to secondary computing devices. A second sequence number corresponding to a second version of the training data may be determined. The second sequence number may be sent to secondary computing devices. Based on inputting the second sequence number into the identical copy of the distributed machine learning model, copies of the second version of the training data may be generated in the secondary computing devices.