Semi-supervised Learning Apparatus for Multi-class Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semi-supervised learning techniques struggle to efficiently train machine learning models that classify objects as any of a plurality of classes, especially when the classes assigned in training data differ across datasets.
Innovation Solution
A learning apparatus and method that generates a machine learning model capable of classifying objects by using data without pre-assigned labels, employing multiple trained models to classify and refine the target model based on their collective classification results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If existing semi-supervised learning techniques are used to train models with few labeled data, then the model can be trained with limited labeled data, but the technique cannot efficiently handle models that classify objects as any of a plurality of classes when class labels differ across datasets
Solution Approach 1:
The patent divides the multi-class classification problem into multiple binary classification sub-problems. For each class, a separate binary classifier is trained to distinguish that class from all other classes. This segmentation allows the system to handle multiple classes efficiently by solving simpler binary problems rather than one complex multi-class problem, directly addressing the contradiction between handling multiple classes and maintaining training efficiency
Solution Approach 2:
The patent creates a universal training framework that can handle any number of classes and any combination of labeled/unlabeled data. The same semi-supervised learning approach is applied uniformly across all binary classifiers, making the system versatile for different multi-class scenarios while maintaining consistent training efficiency through the standardized binary classification approach
2Quantity of substance
If training data with different class labels from various entities is used, then more diverse training data is available, but data sets with incomplete class assignments cannot be used for training
Solution Approach 1:
The patent segments the training data processing into class-specific binary classification tasks. Each binary classifier focuses on one class, allowing the system to selectively use data that contains that class regardless of what other classes are present in the dataset. This enables compatibility across datasets with different class compositions while maximizing the quantity of usable training data
Solution Approach 2:
The patent extracts and utilizes only the relevant portions of diverse training datasets for each binary classification task. For training a binary classifier for class A, only data containing class A labels is extracted and used, while data without class A labels is excluded from that specific training task. This extraction approach enables the system to leverage diverse data from multiple sources while maintaining compatibility by using only the applicable subsets
3Measurement precision
If multiple trained models are used to classify data for training a target model, then classification accuracy can be improved, but the complexity of the training system increases
Solution Approach 1:
The patent segments the complex multi-class classification problem into multiple simple binary classification tasks. Each binary classifier is trained independently on a simplified task, reducing the complexity of individual model training while collectively achieving high multi-class classification accuracy. The segmentation transforms one complex problem into multiple manageable sub-problems
4Quantity of substance
If data without pre-assigned labels is used for training, then more diverse training data can be utilized, but the data cannot be directly used without label assignment
Solution Approach 1:
The patent performs preliminary classification of unlabeled data using the trained binary classifiers before using it for training the target model. The binary classifiers generate predicted labels for the unlabeled data, and these predictions are then used as training labels. This preliminary action transforms unlabeled data into labeled training data, enabling the use of diverse unlabeled datasets while maintaining ease of training through automated label generation
Data Source
AI summary
A learning apparatus which generates a predetermined machine learning model capable of classifying an object included in data to be processed as any of a plurality of classes is disclosed. The apparatus generates the trained predetermined machine learning model, using data to be processed to which a label representing a class of the object is not assigned. The apparatus classifies the data using a trained first machine learning model capable of performing classification regarding a first class among the plurality of classes and classifies the data using a trained second machine learning model capable of performing classification regarding a second class among the plurality of classes, and trains the predetermined machine learning model based on classification results of the predetermined machine learning model, the first machine learning model, and the second machine learning model for the data.