Environment-Adaptive Model Creation Using Basic Model Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems provide learned models that are limited to a specific purpose, leading to inappropriate selections and inaccurate results for user-side devices.
Innovation Solution
A model creation device that determines a basic model adapted for a target environment by analyzing candidate models from different environments, using teacher data to establish a correlation, and creates a target model through fine-tuning or transfer-learning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If a learned model is selected from a database adapted for the purpose of use of the user-side device, then the model can be provided in a short time, but the selected model may not be appropriate and may not provide accurate results
Solution Approach 1:
The patent applies dynamics by making the model selection process adaptive rather than static. The selection unit dynamically selects candidate models based on environmental information from the target environment and training environment, allowing the system to adapt to different contexts while maintaining fast provision time.
Solution Approach 2:
The patent changes the selection parameters from simple purpose-matching to environment-based selection. By using environmental information as the selection criterion, the system can choose models that are not only functionally appropriate but also environmentally adapted, improving accuracy without sacrificing speed.
2Loss of time
If a learned model selected from the database is used without fine-tuning, then the model provision time is short, but the model may not be appropriate for the target environment
Solution Approach 1:
The patent applies partial action by performing selective fine-tuning only when necessary. The selection unit determines whether fine-tuning is needed based on environmental matching, allowing the system to skip fine-tuning for well-matched models (saving time) while applying it only when environmental differences require adaptation.
Solution Approach 2:
The patent uses preliminary action by pre-selecting candidate models that are already adapted to similar environments. This preliminary environmental matching reduces the need for extensive fine-tuning later, as the selected models are already partially adapted to the target environment.
3Measurement precision
If multiple candidate models are evaluated and fine-tuned for the target environment, then the model accuracy is improved, but the complexity of the model creation process increases
Solution Approach 1:
The patent segments the model creation process into distinct stages: candidate model selection based on environmental information, and optional fine-tuning. This segmentation allows the system to handle complexity in a structured way, evaluating multiple candidates systematically and applying fine-tuning only to selected candidates, rather than processing all models through the entire pipeline.
4Ease of operation
If a model is selected based solely on purpose matching, then the selection process is simple, but the model may not be appropriate for the specific target environment
Solution Approach 1:
The patent introduces environmental information as an intermediary between the model database and the selection process. This intermediary layer provides context about both the target and training environments, enabling the selection unit to make more informed decisions that balance simplicity with environmental suitability.
Data Source
AI summary
A device for creating a target model adapted for a target environment and outputting a target output when a target input is input includes: a unit for determining candidate models adapted respectively for candidate environments different from the target environment and outputting a reference output respectively when a reference input is input; a unit for determining a basic environment from among the candidate environments based on the data of the reference output obtained by respectively inputting teacher data of the reference input related to the target environment into the candidate models, and teacher data of the reference output corresponding to the teacher data of the reference input; a unit for determining, as a basic model, a model adapted for the basic environment and outputting the target output when the target input is input; and a unit for creating the target model based on the basic model.