Edge Device Model Training via Server-Side Meta-Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI model training methods require large-scale resources and struggle to optimize training methods for specific tasks on edge devices due to calculation costs and specification limitations, making it difficult to efficiently train models for tasks like image recognition and speech recognition on edge devices without sufficient data or resources.
Innovation Solution
A system that uses meta-learning to select an optimum training method on a server based on task information extracted from edge devices, considering available specifications to enable efficient model training on edge devices, allowing for the use of data sets collected locally while ensuring the selected method can be implemented within the device's capabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If deep learning is performed using enormous amounts of training data sets, then model accuracy is improved, but large-scale operation resources are required
Solution Approach 1:
The patent segments the model training process into two distinct phases: (1) a pre-training phase performed on a server with large-scale resources using enormous training data sets to obtain a base model, and (2) a fine-tuning phase performed on edge devices with limited resources using task-specific data to adapt the base model. This segmentation allows high-accuracy model development to be separated from resource-constrained deployment, resolving the contradiction between model accuracy and operation resource requirements.
2Measurement precision
If a model is trained by using a server or distributed learning, then model accuracy is improved, but the trained model must be mounted on edge devices with limited capabilities
Solution Approach 1:
The patent applies preliminary action by performing the computationally intensive model training and base model creation on a server before deployment to edge devices. The server prepares the base model in advance using large-scale training data, and then this pre-prepared model is transferred to edge devices for fine-tuning with task-specific data. This preliminary preparation enables edge devices with limited capabilities to achieve high model accuracy without performing the entire training process locally.
3Measurement precision
If training data corresponding to a specific task is collected, then model performance for that task is improved, but data shortage remains for certain applications like medical imaging
Solution Approach 1:
The patent merges two data sources to address data shortage: (1) large-scale general training data used on the server to create a base model with broad knowledge, and (2) task-specific data collected locally on edge devices used for fine-tuning. This merging approach allows the model to benefit from both enormous general data and limited task-specific data, resolving the contradiction between achieving task-specific performance and overcoming data shortage for applications like medical imaging.
Data Source
AI summary
The information processing apparatus includes: a management unit that stores a correspondence relationship between a training method for a model and task information of the model; and a selection unit that selects an optimum training method for task information input from a predetermined device and outputs the optimum training method to the device. The management unit associates pieces of specification information necessary for implementing training methods with the training methods, respectively, and stores the pieces of specification information and the training methods. The selection unit selects an optimum training method within a range of a specification available for training in the device.