Ontology-Guided Recognition Training Reducing CNN Complexity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine training recognition techniques face challenges in flexibility and accuracy when adapting to various environments and user needs, as they often require a large number of parameters to be learned again if the initial recognition targets do not cover future needs, leading to inefficiencies and overfitting.
Innovation Solution
A recognition training apparatus that generates relevance between a specific domain and potential recognition targets using ontology information, selects the appropriate targets based on this relevance, and trains a recognizer using tailored training data, thereby reducing the complexity of the recognizer and avoiding overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If pre-training is performed on a large number of categories (e.g., 1000 categories in ImageNet) to cover potential user needs, then the recognizer can be flexibly adapted to various recognition targets, but the device complexity and number of parameters increase enormously
Solution Approach 1:
The patent segments the recognition training process into two distinct phases: pre-training on a large-scale dataset (ImageNet with 1000 categories) to learn general features, and fine-tuning on a smaller, domain-specific dataset to adapt to specific recognition targets. This segmentation allows the system to maintain flexibility while controlling complexity by only training on relevant categories for each specific application.
Solution Approach 2:
The patent performs pre-training in advance on a large-scale dataset to establish a foundation of general recognition capabilities. This preliminary action enables the CNN to learn transferable features that can be later fine-tuned for specific domains, reducing the need to train from scratch and thereby controlling parameter complexity while maintaining adaptability.
2Device complexity
If the pre-training uses a limited number of categories to reduce parameters, then the device complexity is reduced, but the adaptability to cover various user needs deteriorates
Solution Approach 1:
The patent performs pre-training in advance on a large-scale dataset to establish a foundation of general recognition capabilities. This preliminary action enables the CNN to learn transferable features that can be later fine-tuned for specific domains, reducing the need to train from scratch and thereby controlling parameter complexity while maintaining adaptability.
3Measurement precision
If fine-tuning is performed on a domain-specific dataset after pre-training, then the recognition accuracy for specific domains is improved, but the training time and computational resources increase
Solution Approach 1:
The patent performs pre-training in advance on a large-scale dataset to establish a foundation of general recognition capabilities. This preliminary action enables the CNN to learn transferable features that can be later fine-tuned for specific domains, reducing the need to train from scratch and thereby controlling parameter complexity while maintaining adaptability.
Solution Approach 2:
The patent applies partial fine-tuning to the pre-trained CNN model, updating only certain layers or parameters with domain-specific data rather than retraining the entire network. This partial action achieves improved recognition accuracy for specific domains while significantly reducing the training time and computational resources compared to full retraining.
4Measurement precision
If manual selection of recognition targets for pre-training is performed to match user needs, then the recognition accuracy is improved, but the labor and time required for selection increase
Solution Approach 1:
The patent enables the system to automatically adapt to domain-specific requirements through fine-tuning on domain-specific datasets without requiring manual selection of recognition targets. The pre-trained CNN automatically learns and adapts to the specific domain through the fine-tuning process, eliminating the need for manual target selection while maintaining high recognition accuracy.
Data Source
AI summary
A recognition training apparatus according to the present invention includes a generation unit configured to generate relevance between a specific domain and a candidate of a recognition target based on ontology information expressing a concept structure of the specific domain, a selection unit configured to select the recognition target from the candidate of the recognition target based on the relevance generated by the generation unit, and a training unit configured to train a recognizer using training data regarding the recognition target selected by the selection unit.