Momentum Prototypes for Noisy Label Correction in Partially Supervised Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Manual annotation of large datasets for neural networks in computer vision is labor-intensive and expensive, while self-supervised learning methods do not achieve comparable performance, especially with noisy labels from web images.
Innovation Solution
A partially supervised learning mechanism using momentum prototypes that correct noisy labels by minimizing the difference between image embeddings and class prototypes, generating pseudo labels to identify and remove out-of-distribution samples, and updating prototypes based on new training samples.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to train neural networks, then prediction accuracy is improved, but labor cost and time consumption increase significantly
Solution Approach 1:
The system performs self-supervised learning by automatically generating pseudo-labels from web images without requiring manual annotation. The momentum prototype framework enables the model to learn from partially labeled data by comparing image embeddings against learned class prototypes, allowing the system to service itself in terms of data labeling.
Solution Approach 2:
The momentum prototype acts as an intermediary between the image embeddings and the class labels. Instead of directly comparing images to manually annotated labels, the system uses momentum prototypes as a mediating representation that captures class characteristics, enabling automatic learning from web images with noisy or absent labels.
2Productivity
If self-supervised learning is used to avoid manual annotation, then productivity is improved, but prediction accuracy deteriorates
Solution Approach 1:
The system performs preliminary action by pre-training momentum prototypes using available web images before final model training. This preliminary learning phase allows the system to establish class representations in advance, improving subsequent prediction accuracy without requiring manual annotation of the entire training dataset.
Solution Approach 2:
The system implements feedback mechanisms where prediction results are used to update momentum prototypes iteratively. The prototypes are refined based on the distribution of image embeddings, creating a feedback loop that continuously improves prediction accuracy while maintaining automated processing without manual intervention.
3Quantity of substance
If web images with noisy labels are used for training, then data quantity is improved, but reliability deteriorates
Solution Approach 1:
The system extracts reliable class characteristics from web images by comparing image embeddings against momentum prototypes. By taking out only the essential class-relevant features and ignoring noisy label information, the system can utilize large quantities of web images while maintaining training reliability through prototype-based validation.
Data Source
AI summary
A learning mechanism with partially-labeled web images is provided while correcting the noise labels during the learning. Specifically, the mechanism employs a momentum prototype that represents common characteristics of a specific class. One training objective is to minimize the difference between the normalized embedding of a training image sample and the momentum prototype of the corresponding class. Meanwhile, during the training process, the momentum prototype is used to generate a pseudo label for the training image sample, which can then be used to identify and remove out of distribution (OOD) samples to correct the noisy labels from the original partially-labeled training images. The momentum prototype for each class is in turn constantly updated based on the embeddings of new training samples and their pseudo labels.