Data Conversion Learning for Domain Adaptation with Missing Classes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing domain adaptation methods based on adversarial learning struggle when target data does not include all classes present in the source data, leading to decreased accuracy in pattern recognition due to unmatched data distributions.
Innovation Solution
A data conversion learning apparatus that calculates domain and class certainty factors to estimate the likelihood of data belonging to specific domains and classes, and uses these factors to calculate losses for non-appearing classes, thereby performing machine learning to minimize the sum of losses across all classes and domains, including those without target data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If domain adaptation is performed using general adversarial learning approaches, then data distribution matching is achieved, but accuracy decreases for non-appearing classes when target data is incomplete
Solution Approach 1:
The patent segments the loss calculation into two distinct parts: a first loss for appearing classes (where target data exists) and a second loss for non-appearing classes (where target data does not exist). This segmentation allows the model to handle different class scenarios separately, preventing the incomplete target data from negatively impacting the overall domain adaptation accuracy.
Solution Approach 2:
The patent introduces a fictitious target data generator as an intermediary component that synthesizes target domain data for non-appearing classes. This intermediary enables the model to estimate losses for classes without actual target data by generating pseudo-target data, thereby bridging the gap between source and target domain distributions for all classes including non-appearing ones.
2Measurement precision
If all classes in source data are required to appear in target data, then domain adaptation can be performed accurately, but data collection cost increases
Solution Approach 1:
The patent applies partial action by only requiring target data collection for appearing classes rather than all classes. The model intentionally leaves non-appearing classes without target data and handles them through special loss calculation and fictitious data generation, thereby reducing the quantity of target data needed while maintaining acceptable domain adaptation precision.
Solution Approach 2:
The patent uses copying by creating fictitious target data through a fictitious target data generator for non-appearing classes. This synthetic copying allows the model to proceed with domain adaptation without requiring actual target data for all classes, reducing data collection requirements while maintaining the ability to handle all source classes.
3Reliability
If adversarial learning is used to make domain identification difficult, then domain adaptation is achieved, but class identification accuracy decreases for non-appearing classes
Solution Approach 1:
The patent applies local quality by differentiating the loss calculation and data handling strategies for appearing classes versus non-appearing classes. For non-appearing classes, the model uses special loss functions and fictitious target data, while appearing classes use standard loss functions with actual target data. This localized differentiation preserves class identification accuracy for non-appearing classes while maintaining domain adaptation reliability.
Data Source
AI summary
A data conversion learning apparatus includes a data conversion unit that performs data conversion of source data and target data, a first deduction unit that deduces data of a non-appearing class on the basis of a domain certainty factor acquired by a domain identification using converted data, a second deduction unit that deduces data of a non-appearing class on the basis of a class certainty factor acquired by a class identification using converted data, a class identification learning unit that performs machine learning for class identification using the data of the non-appearing class deduced by the first deduction unit and the source data and the target data which are inputs, and a domain identification learning unit that performs machine learning for domain identification using the data of the non-appearing class deduced by the second deduction unit and the source data and the target data which are inputs.