Control Vector Categorization for ML Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems face challenges in categorization due to the quality of incorrect answers, where some incorrect answers are more or less correct based on relationships between categories, and existing methods do not effectively utilize these relationships to improve categorization accuracy.
Innovation Solution
The use of a control vector with three different values representing a matching category, a non-matching category, and a related category allows the machine learning system to differentiate between near misses and unrelated categories, weighting incorrect answers based on their similarity to the correct answer, thereby improving the quality of categorization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional binary classification (match/non-match) is used, then the system is simple to implement, but it cannot differentiate between near misses and unrelated categories
Solution Approach 1:
The patent segments the traditional binary classification into three distinct categories represented by different values in the control vector: exact match (first value), related category (second value), and unrelated category (third value). This segmentation allows the system to differentiate between near misses and completely unrelated categories, improving measurement precision without excessive complexity
Solution Approach 2:
The patent applies local quality by assigning different values to different positions in the control vector based on the relationship between the predicted category and the actual category. Each position in the control vector has a specific quality (value) that reflects the degree of match, allowing the system to provide localized, nuanced feedback for each category comparison
2Reliability
If all incorrect answers are treated equally, then the training process is simple, but the system cannot learn from the quality of incorrect answers
Solution Approach 1:
The patent implements feedback by using the control vector values to guide the training process. The loss function uses these values to provide differentiated feedback: exact matches receive strong positive feedback, related categories receive moderate feedback, and unrelated categories receive weak or negative feedback. This allows the system to learn from the quality of incorrect answers and improve reliability
Solution Approach 2:
The patent applies asymmetry by treating different types of incorrect answers differently in the training process. The control vector assigns asymmetric weights: first value for exact matches, second value for related categories, and third value for unrelated categories. This asymmetric treatment preserves information about the quality of incorrect answers, allowing the system to learn more effectively
Data Source
AI summary
A method of training a machine learning system, the method including: training the machine learning system in category identification against a test case wherein the machine learning system outputs an answer vector and the answer vector is compared against a control vector where the control vector comprising three different values, the values comprising: a first value for a matching category for the test case; a second value for a non-matching category for the test case; and a third value for a first category related to the matching category for the test case wherein the third value differs from the first value differs from the second value.