Pattern Recognition Dictionary Correction for Noise Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing pattern recognition methods face challenges in maintaining high classification precision when data used for learning contains noise or has a small number of data points, leading to poor performance in recognizing new, unlearned data.
Innovation Solution
A pattern recognition system that calculates losses for each class and differences between classes to create a recognition dictionary, which is then corrected to minimize total losses and differences, enabling improved classification of new data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a typical pattern recognizer uses soft margin classification to maximize margin between classes, then classification precision for clean data is improved, but classification precision for new data deteriorates when learning data contains noise or is limited in quantity
Solution Approach 1:
The patent segments the loss calculation into class-specific components, calculating loss for each class separately rather than treating all classes uniformly. This allows the system to identify and handle noisy or limited data on a per-class basis, improving reliability for new data while maintaining precision for clean data.
Solution Approach 2:
The patent introduces a novel loss function that changes the parameter optimization approach by minimizing the sum of losses for each class and the sum of loss differences between classes. This parameter change in the objective function enables the classifier to adapt better to noisy or limited learning data, improving generalization to new data.
2Device complexity
If the recognition dictionary is created to minimize total loss without considering class-specific loss differences, then computational simplicity is maintained, but classification precision for new data deteriorates
Solution Approach 1:
The patent segments the loss calculation into class-specific components, calculating loss for each class separately. This segmentation adds computational detail that improves classification precision for new data while maintaining a structured approach that doesn't excessively increase complexity.
Solution Approach 2:
The patent incorporates loss difference calculation between classes as a feedback mechanism. By computing how losses differ between classes and using this information to adjust the recognition dictionary, the system achieves better precision for new data without requiring complex external validation processes.
3Manufacturing precision
If margin boundaries are maximized to separate classes clearly, then separation accuracy is improved, but robustness to noise in learning data deteriorates
Solution Approach 1:
The patent changes the optimization parameter from单纯 maximizing margin to minimizing a composite loss function that includes both class-specific losses and loss differences between classes. This parameter change makes the margin boundaries more robust to noise while maintaining clear separation accuracy.
Solution Approach 2:
The patent converts the harmful effect of noise and limited data into a beneficial signal by using loss difference calculation. The differences in loss between classes provide information about data quality and class separability, allowing the system to adjust margins appropriately even in the presence of noise.
Data Source
AI summary
A loss calculation element calculates a loss of a feature vector for each class, the feature vector indicating a feature of data belonging to the class. A loss difference calculation element calculates a sum of differences of losses between the classes based on the loss calculated for each class. A recognition dictionary creation element creates a recognition dictionary based on a total sum of losses calculated for the classes and the total sum of loss differences between the classes. The recognition dictionary creation element corrects, on a basis of an input feature vector, a recognition dictionary so as to minimize a sum of the total sum of losses for the classes calculated by the loss calculation element and the total sum of loss differences between the classes calculated by the loss difference calculation element, and the pattern classification element classifies a pattern of data using the corrected recognition dictionary.