Domain Adaptation via Pseudo-Label Aggregation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing domain adaptation methods struggle when adapting a model from a source domain to a target domain without access to training data or detailed implementation information from the source domain, and when the target domain data is unlabeled.
Innovation Solution
The method involves classifying training data from the target domain using a group of source models from the source domain, generating pseudo labels by aggregating classification results based on source model weights, and training a target model using these pseudo labels and target domain data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If domain adaptation is performed without access to source domain training data or detailed model information, then model adaptability to target domain is improved, but knowledge transfer accuracy deteriorates
Solution Approach 1:
The patent introduces pseudo-labels as an intermediary mechanism to bridge the gap between source models and target domain data. The source models generate predictions on target domain data, these predictions are aggregated through voting to create pseudo-labels, which then serve as training labels for the target model. This intermediary approach enables knowledge transfer without direct access to source training data, resolving the contradiction between adaptability and accuracy.
Solution Approach 2:
The patent creates a copy of the labeling function by using source models to generate pseudo-labels for target domain data. Instead of directly transferring source model parameters or accessing source training data, the system copies the functional outcome (classifications) and uses aggregated voting results to create synthetic labels that mimic the structure and quality of true labels, enabling effective target model training.
2Reliability
If pseudo labels are generated by aggregating classification results from multiple source models, then classification reliability in target domain is improved, but computational complexity increases
Solution Approach 1:
The patent merges the classification results from multiple source models through a voting aggregation mechanism. Each source model independently classifies target domain data, and the system combines these predictions by counting votes for each class. This merging approach improves classification reliability by leveraging diverse model perspectives while maintaining computational efficiency through simple majority voting rather than complex ensemble methods.
3Quantity of substance
If target domain data is unlabeled, then data availability for training is improved, but training effectiveness deteriorates
Solution Approach 1:
The patent implements a self-service labeling mechanism where the source models automatically generate pseudo-labels for the target domain data without human intervention. The system uses the source models' predictions, aggregates them through voting, and directly creates training labels from these aggregated results. This self-service approach transforms unlabeled data into effectively labeled data, enabling automated training pipeline that maintains both data availability and training effectiveness.
Data Source
AI summary
A computer implement method for domain adaptation. According to the method, training data from a target domain may be classified by using a group of source models of a source domain, where each of the source models is trained to classify data from the source domain. A first pseudo label indicating a category of the training data may be generated by aggregating classification results of the source models based on respective weights of the source models on predetermined data categories. A target model of the target domain may be trained based on the training data and the first pseudo label and the second pseudo label, where the target model is trained to classify data from the target domain.