Semi-supervised Label Generation via Dual Model Consensus
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for appearance inspection using machine learning models face challenges in generating reliable correct answer labels efficiently, especially in semi-supervised learning and domain adaptation scenarios, leading to low inference accuracy and increased costs due to the time and effort required for data collection and label generation.
Innovation Solution
A method involving the use of two trained inference models, one from the source domain and another from a domain with added disturbance, to generate a third correct answer label based on the match between their inference results, allowing for automated and cost-effective label generation with improved reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If multiple data sets are collected to improve inspection accuracy, then the accuracy of appearance inspection is improved, but the time and effort required for data collection and label generation increases
Solution Approach 1:
The system uses the trained machine learning model itself to generate pseudo correct answer labels for new training data. The model performs self-labeling by automatically inferring labels for unlabeled data, eliminating the need for manual annotation and enabling continuous expansion of the training data set without proportional increase in time investment
Solution Approach 2:
The system pre-generates pseudo correct answer labels using the trained model before actual inspection tasks. By preparing labeled training data in advance through automated pseudo-labeling, the system reduces the time required for data preparation and enables faster deployment of inspection models
2Reliability
If manual generation of correct answer labels is performed to ensure label accuracy, then the reliability of training data is improved, but the cost and time required for label generation increases
Solution Approach 1:
The trained machine learning model automatically generates pseudo correct answer labels for new training data without human intervention. This self-labeling capability maintains label consistency with the model's learned patterns while dramatically increasing label generation throughput and reducing costs
Solution Approach 2:
The system copies the labeling pattern learned from manually annotated data and applies it automatically to new data through the trained model. This copying of labeling knowledge from labeled to unlabeled data maintains reliability while scaling production
3Measurement precision
If domain adaptation is performed to improve inference accuracy for target domain data, then the inference accuracy is improved, but the complexity of the learning process increases
Solution Approach 1:
The system adapts to domain differences by generating pseudo labels that reflect the target domain characteristics. By changing the labeling parameters through the trained model's learned representations rather than manually adjusting model architecture or performing complex domain adaptation, the system achieves domain adaptation with reduced complexity
Data Source
AI summary
A label generation method according to one aspect of the present invention prepares a first inference model trained on a first dataset obtained from a source domain, and a second inference model trained on a second dataset including second training data generated by adding a disturbance to first training data, and generates a third correct answer label for third training data, on the basis of a consensus of the prepared trained first inference model and second inference model.