Modular MoE Model Selection for Lower-VRAM Domain AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional MoE architectures face challenges with high VRAM requirements, inefficient resource utilization, and the need for one-size-fits-all AI models, limiting their performance in on-device environments.
Innovation Solution
A method and system for specifying a model based on a MoE architecture that involves modularizing specialized models, using a router to allocate tasks efficiently, and constructing a DMoE model for optimized data processing.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional MoE architectures are used to process specific tasks based on large-volume data, then computational efficiency is improved, but VRAM requirements increase significantly
Solution Approach 1:
The patent segments the MoE model into specialized module models that are independently trained and stored. Each specialized model handles specific tasks or data types, allowing the system to load only the necessary segments rather than maintaining the entire MoE architecture in memory simultaneously.
Solution Approach 2:
The patent extracts and separates the specialized models from the conventional MoE architecture, storing them as independent discrete models. This extraction allows the system to selectively activate and load only the required specialized models based on the input data characteristics, reducing the VRAM burden.
2Productivity
If conventional MoE architectures are used to manage large-sized models, then overall efficiency is enhanced, but resource utilization becomes inefficient
Solution Approach 1:
The patent implements a dynamic model selection mechanism where the system automatically determines which specialized module model to activate based on the characteristics of the input data. This dynamic approach ensures that computational resources are allocated efficiently by activating only the necessary models for each specific task, rather than maintaining all models in an active state.
3Ease of operation
If one-size-fits-all AI models are used to provide services, then model deployment is simplified, but AI analysis performance tailored to specific contexts deteriorates
Solution Approach 1:
The patent applies local quality by creating specialized module models that are optimized for specific domains, data types, or tasks. Each specialized model has tailored parameters and architectures suited to its specific function, allowing the system to achieve high precision for context-specific analysis while maintaining a unified deployment framework through the router mechanism.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method for specifying a model based on an artificial intelligence model including a plurality of models according to an embodiment of the present invention comprises: performing MoE (Mixture of Experts) architecture-based training based on an artificial intelligence model which is a combination of a plurality of specialized models and a router (gating network); obtaining specialized model feature information which is information that specifies features of the specialized models based on the performed MoE architecture-based training; generating a specialized module model which is a specialized model that is matched with the obtained specialized model feature information and is independent and discrete; obtaining given domain information; detecting specialized model feature information of a feature corresponding to the obtained domain information; extracting a specialized module model that is matched with the detected specialized model feature information; constructing a DMoE model which is an artificial intelligence model which is a combination of the extracted specialized module model and the router; and providing output data for given input data based on the constructed DMoE model.