Data Augmentation and Label Propagation for Unlabeled Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning methods, such as active learning and self-supervised learning, require significant human effort for labeling datasets, and label propagation is limited by the need for high-confidence data, making it difficult to efficiently reduce human costs for labeling, especially in cases where unlabeled data is distant from labeled data.
Innovation Solution
The approach involves data augmentation and label propagation to provide weak labels to unlabeled data, using a data gathering apparatus that includes a CPU, memory, display controller, and external recording medium processor to perform data augmentation and label propagation, and a camera to obtain consecutive unlabeled data close to labeled data, thereby reducing human labeling costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual labeling is used to provide correct answer labels to datasets, then labeling accuracy is improved, but labeling cost increases
Solution Approach 1:
The system performs self-labeling by automatically generating pseudo-labels for unlabeled data through data augmentation and label propagation. The learning model labels its own training data by propagating labels from labeled data to augmented unlabeled data, eliminating the need for manual human labeling while maintaining acceptable accuracy through iterative refinement.
Solution Approach 2:
Data augmentation serves as an intermediary technique that transforms existing labeled data into multiple augmented versions. These augmented data pieces act as intermediaries to propagate labels to new unlabeled data, bridging the gap between limited labeled data and large volumes of unlabeled data without requiring direct manual labeling of all data.
2Quantity of substance
If self-supervised learning with label propagation is used, then labeling cost is reduced, but labeling reliability deteriorates when data is distant from labeled data
Solution Approach 1:
The system performs preliminary data augmentation on existing labeled data before attempting label propagation. By pre-processing labeled data through multiple augmentation techniques to create a robust set of augmented samples, the system prepares a stronger foundation for label propagation, improving reliability when propagating labels to distant unlabeled data points.
Solution Approach 2:
The system applies extensive data augmentation beyond what would be minimally required, generating a large number of augmented data pieces from each labeled sample. This excessive action creates multiple intermediate representations that improve the chances of finding reliable propagation paths to distant unlabeled data, compensating for the distance through multiple transformation steps.
3Productivity
If active learning with confidence level filtering is used, then labeling efficiency is improved, but productivity decreases when handling low confidence data
Solution Approach 1:
The system automatically handles low-confidence data through self-labeling mechanisms rather than requiring manual intervention. When the learning model encounters unlabeled data, it generates pseudo-labels through data augmentation and label propagation, automatically resolving uncertainties without human time investment, thus maintaining high efficiency even with large volumes of low-confidence data.
Solution Approach 2:
Data augmentation creates intermediate augmented samples that serve as mediators between labeled and unlabeled data. These intermediate representations provide confidence bridges, allowing the system to propagate labels through multiple steps rather than requiring direct high-confidence matches, thereby reducing the need for manual handling of low-confidence cases.
4Adaptability or versatility
If data augmentation is performed on unlabeled data, then labeling coverage is improved, but device complexity increases
Solution Approach 1:
The data augmentation process is segmented into discrete, modular operations that can be applied independently to labeled data. By breaking down augmentation into separate transformation steps (rotation, scaling, cropping, etc.), the system manages complexity through modularity while achieving comprehensive labeling coverage through systematic application of multiple segmentation operations.
Data Source
AI summary
A non-transitory computer-readable recording medium has stored therein a data gathering program executable by one or more computers, the data gathering program including: performing data augmentation on unlabeled data; providing a specification label to a group of augmented data pieces generated by the data augmentation, the specification label indicating that labels of the augmented data pieces all match; and providing, when a label for one data piece of the augmented data pieces is determined, the label to one or more data pieces each provided with a specification label that is same as a specification label of the one data piece.