Neural Network Training with Partial Labels for Multi-Class Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods face challenges in constructing neural networks that accurately classify images into multiple classes due to the high burden of creating large amounts of training data with precise correct answer labels, particularly when classifying regions like the liver and lung lobes in medical images or other expression media.
Innovation Solution
A learning device and method that integrate probabilities from partial correct answer labels to train neural networks, allowing classification into multiple classes using reduced training data by integrating probabilities based on available labels and minimizing the need for complete labeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complete correct answer labels for all classes are prepared for training data, then classification accuracy into multiple classes is improved, but the burden on data creators and time required for data preparation increases significantly
Solution Approach 1:
The patent applies partial action by training the neural network using only partial correct answer labels (labels for one or more specific classes) rather than requiring complete labels for all classes. The system integrates probabilities from multiple training datasets, each containing labels for different classes, to achieve comprehensive multi-class classification without the need for all classes to be labeled in every training sample.
2Measurement precision
If complete correct answer labels for all classes are prepared for training data, then classification accuracy into multiple classes is improved, but the complexity of data creation process increases
Solution Approach 1:
The patent segments the data creation process by dividing the training data into multiple separate datasets, where each dataset contains correct answer labels for specific classes rather than requiring all classes to be labeled together. This segmentation reduces the complexity of the data creation process while still enabling comprehensive multi-class classification through probability integration.
3Measurement precision
If separate trained neural networks are prepared for each class, then classification accuracy for each class is improved, but the system complexity and number of models required increases
Solution Approach 1:
The patent merges multiple training datasets with different label configurations into a single unified neural network model. By integrating probabilities from multiple datasets during training, the system achieves per-class classification accuracy comparable to separate models while maintaining a single model architecture, thus reducing system complexity.
4Adaptability or versatility
If training data with multiple class labels is prepared, then multi-class classification capability is improved, but the amount of required training data increases
Solution Approach 1:
The patent creates training datasets that serve multiple functions: each dataset contains labels for specific classes and can be used to train the neural network for those classes. By integrating probabilities from multiple such datasets, the system achieves comprehensive multi-class classification capability without requiring a single large dataset containing all classes, thus reducing the total training data volume required.
Data Source
AI summary
A processor is configured to: acquire training data that consists of a learning expression medium and a correct answer label for at least one of a plurality of types of classes included in the learning expression medium; input the learning expression medium to a neural network such that probabilities that each class included in the learning expression medium will be each of the plurality of types of classes are output; integrate the probabilities that each class will be each of the plurality of types of classes on the basis of classes classified by the correct answer label of the training data; and train the neural network on the basis of a loss derived from the integrated probability and the correct answer label of the training data.