Multi-Label Learning Model Using Dual-Loss Feature Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Preparing training data for multi-label data is particularly time-consuming due to numerous label combinations, and existing methods are not effective in increasing the accuracy of learning models using a small amount of training data.
Innovation Solution
A learning system that includes first and second calculation units to calculate losses based on the output and feature amounts of multi-label query and support data, with an adjustment unit to refine the model parameters using these losses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a larger amount of training data is prepared, then the accuracy of the learning model can be increased, but the time required to prepare training data increases significantly
Solution Approach 1:
The patent segments the training process into two distinct loss calculation components: a first loss based on output accuracy and a second loss based on feature amount similarity. This segmentation allows the system to optimize model accuracy through multiple mechanisms simultaneously, reducing the need for extensive training data while maintaining high accuracy performance
Solution Approach 2:
The patent introduces a feature amount parameter that captures semantic similarity between data points. By changing the optimization parameters to include both output loss and feature amount loss, the system can achieve high accuracy with smaller training datasets, as the feature amount parameter helps generalize across similar data points
2Adaptability or versatility
If existing machine learning methods are used for multi-label data, then the method can be applied to single-label data, but it becomes inapplicable or inefficient for multi-label data with many label combinations
Solution Approach 1:
The patent creates a universal learning framework that handles both single-label and multi-label data through the same architecture. The dual-loss function (output loss + feature amount loss) works generally for any label configuration, making the system adaptable to multi-label data without requiring separate methods for different data types
Solution Approach 2:
The feature amount calculation acts as an intermediary mechanism that bridges the gap between raw input data and label classification. By computing feature amounts that capture semantic similarity, the system can handle the complexity of multi-label combinations through a standardized process, reducing the effective complexity of the classification task
Data Source
Figure 1
Figure 2
Figure 3
AI summary
First calculation means (302) of a learning system (S) calculates, when multi-label query data is input to a learning model, a first loss based on an output of the learning model and a target output. Feature amount acquisition means (303) acquires a feature amount of the query data and a feature amount of support data corresponding to the query data, which are calculated based on a parameter of the learning model. Second calculation means (304) calculates a second loss based on the feature amount of the query data and the feature amount of the support data. Adjustment means (305) adjusts the parameter based on the first loss and the second loss.