Training Image Selection Using Parameter Similarity Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image-for-training selection techniques for machine learning models are inadequate, particularly when incorrect user-set categories and imbalanced training data lead to inappropriate data selection, affecting the model's training efficiency and accuracy.
Innovation Solution
An apparatus and method utilizing contrastive learning to train a first machine learning model with a first layer group that generates features, and a second model with a connected second layer group for classification, calculating similarities between parameters before and after training to determine inappropriate images-for-training based on similarity thresholds.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If training data is selected based on user-set correct categories, then the selection process is simple, but the selection becomes unreliable when users set incorrect categories or when skill level affects category setting accuracy
Solution Approach 1:
The patent implements a feedback mechanism where the model's determination results are compared with user-set correct categories, and the difference is used to automatically adjust training data selection. This feedback loop allows the system to learn from user inputs while compensating for potential errors in category setting, thereby improving selection reliability without complicating the user interface.
Solution Approach 2:
The system performs self-service by automatically selecting and adjusting training data based on the difference between model determinations and user-set categories. The selection apparatus autonomously identifies inappropriate training data and adjusts the selection without requiring manual intervention, maintaining operational simplicity while improving reliability through automated correction.
2Productivity
If training data is selected without considering balance and comprehensiveness, then the selection process is faster and simpler, but the model training effectiveness deteriorates due to imbalanced data
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing difference values between model determination results and correct categories for all training data before the selection process. This pre-processing enables the system to quickly identify and select appropriate training data without requiring complex real-time analysis, thus maintaining high selection speed while ensuring data balance and comprehensiveness for effective training.
3Measurement precision
If contrastive learning is used to train the first machine learning model, then feature extraction capability is improved, but the complexity of the training process increases
Solution Approach 1:
The patent segments the training process into two distinct stages: first training the feature extraction model (first machine learning model) using contrastive learning, then training the classification model (second machine learning model) using the extracted features. This segmentation allows the complex contrastive learning to be isolated in the first stage, improving feature extraction capability while managing overall process complexity through modular training architecture.
Data Source
AI summary
An image-for-training selecting apparatus for suitably selecting an image-for-training for training a machine learning model includes at least one processor executing: a first training process of training, by contrastive learning using an images-for-training set, a first machine learning model including a first layer group; a second training process of training a second machine learning model including the first layer group and a second layer group and employing the first machine learning model as a pre-trained model; a first calculating process of calculating a first similarity between a parameter of the first layer group after training by the first training process but before training by the second training process and a parameter of the first layer group after training by the second training process; and a first determining process of determining, based on the first similarity, whether the images-for-training set includes an inappropriate image-for-training.