Large Model Training With Segmented Edge Computing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training large AI models is computationally intensive, time-consuming, and costly, necessitating resource and time-efficient techniques.
Innovation Solution
Training large models using edge computing devices by splitting them into individually trainable components and data chunks, executing these chunks on edge devices, and aggregating the results to form an updated model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional centralized training methods are used for large AI models, then training can be performed with sufficient computational power, but the process becomes computationally intensive, time-consuming, and costly
Solution Approach 1:
The patent divides the large AI model into multiple smaller model components that can be independently trained on edge devices. Each edge device trains a specific component using local data, and the results are aggregated to form the complete trained model. This segmentation enables distributed training across multiple devices, reducing the computational burden on any single device while maintaining overall training efficiency.
Solution Approach 2:
The patent transitions from centralized training to distributed training by utilizing the edge computing dimension. Instead of concentrating all training computations in one location, the system leverages the computational capabilities of multiple edge devices across different locations, effectively adding a spatial dimension to the training process and distributing the computational workload.
2Productivity
If large models are trained using traditional methods, then comprehensive training can be achieved, but the process is time-consuming
Solution Approach 1:
By segmenting the model into smaller components that can be trained simultaneously on different edge devices, the patent enables parallel training operations. Multiple model components are trained concurrently across distributed edge devices, significantly reducing the total training time compared to sequential centralized training while maintaining comprehensive model training.
Solution Approach 2:
The patent implements continuous training where edge devices continuously process and train model components in the background, and the training controller continuously aggregates results. This continuous distributed training operation eliminates idle time and ensures uninterrupted progress, accelerating the overall training process compared to traditional batch processing methods.
3Productivity
If large models are trained using traditional methods, then full training can be performed, but the cost increases significantly
Solution Approach 1:
The patent segments the training workload across multiple edge devices, each contributing to training a specific model component. This distribution reduces the computational cost burden on any single organization or device while maintaining the effectiveness of training the complete large model through aggregation of results from multiple sources.
Solution Approach 2:
The system enables edge devices to perform self-training of model components using their own local data and computational resources. This self-service capability allows distributed participation in the training process, reducing the need for expensive centralized training infrastructure and lowering overall training costs while maintaining training effectiveness.
Data Source
AI summary
An apparatus performs a method for training a large model using edge computing devices. The method includes generating one or more training code chunks using a large model, the training code chunks including a component of the large model and a tuning code for the component. The component is individually trainable; generating multiple training data chunks from training data, the training data chunks capable of being processed by the training code chunks on edge nodes, to train the component in the training code chunk; generating a chunk pair including the training code chunk and the training data chunk; sending the chunk pair to an edge node remote to the training controller; receiving a first processed training code chunk from the edge node; and aggregating the first processed training code chunk with at least one second processed training code chunk to generate an updated large model.