Neural Network Training Ignorable Label Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current neural network training methods using backpropagation algorithms face challenges in accurately classifying data due to incorrect label evaluations, leading to potential training based on incomplete or incorrect data, which can degrade recognition performance.
Innovation Solution
A training system that introduces an 'ignorable evaluation' in addition to 'positive' and 'negative' evaluations, allowing only correct evaluations to be used for training, by adjusting weight coefficients to ignore 'ignorable' labels, thus preventing incorrect training and improving recognition accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If label evaluation is automatically estimated by another training, then workload on annotator is reduced, but correct evaluation cannot be necessarily estimated leading to training based on incorrect evaluation
Solution Approach 1:
The patent introduces an intermediary evaluation status called 'ignorable evaluation' that acts as a mediator between positive and negative evaluations. This intermediary state allows the system to handle uncertain or unverified label evaluations without automatically treating them as correct, thus maintaining evaluation reliability while reducing annotator workload through automated processing of clearly defined cases.
Solution Approach 2:
The patent changes the parameter space of label evaluations by adding a third state 'ignorable evaluation' alongside traditional positive and negative evaluations. This parameter expansion allows the system to distinguish between evaluated and unverified labels, enabling automated processing to proceed without compromising the reliability of training data by excluding uncertain evaluations from influencing model training.
2Quantity of substance
If training data includes incomplete label evaluations, then data quantity is increased, but training accuracy degrades due to incorrect evaluations
Solution Approach 1:
The patent applies local quality by treating different label evaluations differently within the training data. Labels with 'ignorable evaluation' status are handled locally by excluding them from gradient computation and weight updates, while labels with positive or negative evaluations continue to influence training normally. This localized differentiation allows the system to incorporate more training data without uniformly degrading training accuracy.
Solution Approach 2:
The patent segments the training data based on evaluation completeness and reliability. By dividing training samples into those with reliable evaluations (positive/negative) and those with incomplete evaluations (ignorable), the system can process segments differently - using reliable segments for accurate gradient computation while excluding or separately handling incomplete segments, thus maintaining overall training accuracy while increasing usable data quantity.
3Productivity
If backpropagation uses all label evaluations, then training completeness is improved, but incorrect evaluations negatively affect weight coefficient adjustment
Solution Approach 1:
The patent applies partial action by selectively including only reliable label evaluations (positive and negative) in the backpropagation weight adjustment process, while excluding incomplete evaluations (ignorable). This partial inclusion ensures that weight coefficients are adjusted based solely on verified correct evaluations, maintaining reliability of training while still achieving comprehensive training coverage through multiple iterations with reliable data.
Data Source
Figure 1(A)~1(B)
Figure 2
Figure 3(A)~3(B)
AI summary
A training system comprises a training device, and a training data creation device for the training device. The training device trains a neural network by means of a backpropagation algorithm. The training data creation device acquires any one of a positive evaluation indicating that content of the input data coincides with the label, a negative evaluation indicating that content of the input data does not coincide with the label, and an ignorable evaluation indicating exclusion from a training target label, for each label regarding the input data to create training data. In training the neural network for training, the training system adjusts the weight coefficient for the intermediate layer such that the recognition score of the label with the positive evaluation or the negative evaluation comes closer to the ground-truth score of the positive evaluation or the negative evaluation, and makes the recognition score of the label with the ignorable evaluation not affect the adjustment of the weight coefficient for the intermediate layer.