Pseudo-Label Evaluation for Domain Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing techniques for adding pseudo-labels to unlabeled data in information processing systems face challenges in accuracy and efficiency, particularly in regression problems, where the quality of pseudo-labels is not consistently high, leading to increased labeling costs and potential accuracy reductions due to domain differences between labeled and unlabeled data.
Innovation Solution
An information processing system that includes a data input unit for labeled and unlabeled data, a pseudo-label addition unit using a teacher model learned from labeled data, a pseudo-label evaluation unit using an evaluation model to assess and filter pseudo-labels based on a predetermined criterion, and a student model learning unit that utilizes high-quality pseudo-labels to improve model accuracy, along with an optional domain conversion unit to align data domains.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If pseudo-labels are added to unlabeled data using a teacher model, then the volume of usable data increases, but the accuracy of pseudo-labels decreases due to domain differences
Solution Approach 1:
The patent introduces a domain adapter as an intermediary component between the teacher model and the unlabeled data. The domain adapter learns to map features from the source domain (labeled data) to the target domain (unlabeled data), thereby mediating the domain difference and enabling accurate pseudo-label generation without requiring direct training on target domain data.
Solution Approach 2:
The patent changes the parameter space by introducing domain-specific adaptation layers that modify the feature representations. The domain adapter adjusts the feature parameters to align with the target domain distribution, allowing the teacher model to generate accurate pseudo-labels for unlabeled target domain data.
2Ease of manufacture
If pseudo-labels are added to unlabeled data, then the labeling cost decreases, but the reliability of the labeling process worsens due to poor pseudo-label quality
Solution Approach 1:
The patent implements a feedback mechanism where the pseudo-labeled data is used to retrain the domain adapter and teacher model. This iterative feedback loop allows the system to continuously improve pseudo-label quality by learning from previously generated labels, thereby increasing reliability while maintaining low labeling costs.
Solution Approach 2:
The patent performs preliminary domain adaptation training before pseudo-label generation. The domain adapter is pre-trained on labeled source domain data to learn domain-specific transformations, which then enable reliable pseudo-label generation on unlabeled target domain data without requiring expensive manual labeling.
3Measurement precision
If domain differences between labeled and unlabeled data are addressed, then the accuracy is maintained, but the device complexity increases
Solution Approach 1:
The patent segments the system into distinct functional components: a domain adapter module, a teacher model module, and a student model module. Each component handles a specific task (domain adaptation, pseudo-label generation, and final learning), which simplifies the overall system architecture while maintaining accuracy through specialized processing.
Data Source
AI summary
An information processing system includes: a data input unit that inputs labeled data and unlabeled data; a pseudo-label addition unit that adds a pseudo-label to the unlabeled data, by using a teacher model learned by using the labeled data; a pseudo-label evaluation unit that evaluates the pseudo-label, by using an evaluation model learned by using at least one of the labeled data and the unlabeled data, and outputs the pseudo-label that reaches a predetermined evaluation criterion, as an evaluation pseudo-label; a student model learning unit that learns a student model, by using the labeled data, and pseudo-labeled data in which the evaluation pseudo-label is added; and a model output unit that outputs the learned student model.