Semi-Supervised Model Training With One-Bit Label Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning models require a significant amount of labeled data to enhance generalization capability and improve prediction accuracy, and there is a need for a method to enhance prediction accuracy using a smaller amount of labeled data.
Innovation Solution
A training method for a semi-supervised learning model that involves predicting classification categories for unlabeled samples, performing one-bit labeling based on the predictions, and reconstructing the training set to update the model, using different loss functions for different sample sets to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large amount of labeled data is used for training, then the generalization capability and prediction accuracy of the semi-supervised learning model are significantly enhanced, but the cost and time required for data labeling increase substantially
Solution Approach 1:
The patent implements a feedback mechanism where the model's prediction results are fed back into the training process. Unlabeled samples that are correctly predicted are added to the labeled training set in subsequent training iterations, creating a positive feedback loop that continuously improves model accuracy while requiring minimal initial labeled data
Solution Approach 2:
The system performs self-labeling by using its own predictions to generate training labels. The model automatically identifies confidently predicted unlabeled samples and uses its own predictions as labels for retraining, eliminating the need for extensive manual labeling while maintaining high prediction accuracy
2Reliability
If manual labeling is performed extensively to obtain correct labels for unlabeled samples, then the training set quality improves, but the time and labor costs increase significantly
Solution Approach 1:
The system automatically generates labels for unlabeled samples using its own prediction capabilities. The model identifies samples with high prediction confidence and uses its predicted labels directly, eliminating the need for manual annotator intervention and significantly reducing labeling time while maintaining training set quality
Solution Approach 2:
The patent introduces a confidence threshold mechanism as an intermediary between the model's predictions and the labeling process. Samples above the threshold are automatically labeled without manual intervention, while only ambiguous cases require human review, thus reducing overall labeling time while preserving training set reliability
3Measurement precision
If the model is trained iteratively with reconstructed training sets, then the prediction accuracy improves progressively, but the computational complexity and training time increase
Solution Approach 1:
The patent divides the training process into distinct phases: initial training with limited labeled data, iterative refinement phases where predicted samples are added to the training set, and evaluation phases. This segmentation allows the complex training process to be managed in manageable stages while achieving progressive accuracy improvement
Solution Approach 2:
The training process is made dynamic by adaptively adjusting the training set composition in each iteration. The system dynamically adds high-confidence predicted samples to the training set and removes low-confidence samples, creating a living training process that evolves with each iteration rather than following a fixed predetermined structure
Data Source
AI summary
Embodiments of this application disclose a training method for a semi-supervised learning model which can be applied to computer vision in the field of artificial intelligence. The method includes: first predicting classification categories of some unlabeled samples by using a trained first semi-supervised learning model, to obtain a prediction label; and determining whether each prediction label is correct in a one-bit labeling manner, and if prediction is correct, obtaining a correct label (a positive label) of the sample, or if prediction is incorrect, excluding an incorrect label (a negative label) of the sample. Then, in a next training phase, a training set (a first training set) is reconstructed based on the information, and an initial semi-supervised learning model is retrained based on the first training set, to improve prediction accuracy of the model. In one-bit labeling, an annotator only needs to answer “yes” or “no” for the prediction label.