Teaching Data Correction via Model Feedback Loop
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating teaching data for machine learning rely on individual skills, leading to inconsistent annotation and suboptimal identification accuracy, as they do not inherently correct the data to suit the learning model's needs.
Innovation Solution
A method and device that set a correction candidate area for the object area in training images, using an output machine learned from identification or regression results to update the teaching data based on accuracy, thereby improving identification or regression accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If teaching data is generated through manual annotation methods, then the annotation work can be completed, but the identification accuracy is suboptimal due to individual skill variations and lack of systematic correction
Solution Approach 1:
The patent implements a feedback mechanism where the learning model's identification results are used to automatically correct teaching data. The system evaluates identification accuracy and uses this feedback to iteratively improve the teaching data, eliminating manual annotation inconsistencies and systematically enhancing both reliability and measurement precision.
Solution Approach 2:
The system enables self-service correction where the learning model automatically identifies and corrects its own teaching data without external manual intervention. The model uses its own identification results to generate correction information, creating a self-improving cycle that resolves annotation inconsistencies and improves accuracy autonomously.
2Measurement precision
If manual correction of teaching data is performed to improve identification accuracy, then the accuracy can be enhanced, but the workload and time consumption increase significantly
Solution Approach 1:
The patent replaces the mechanical manual correction process with an automated computational system. The learning model automatically generates correction information based on its identification results, substituting human manual labor with algorithmic processing that achieves the same accuracy improvement without the associated time loss.
Solution Approach 2:
The system performs self-correction where the learning model autonomously identifies errors in teaching data and generates correction information without external intervention. This self-service mechanism eliminates the need for time-consuming manual correction while maintaining high identification accuracy.
3Ease of manufacture
If teaching data is generated without systematic correction, then the generation process is simple, but the teaching data does not necessarily indicate the suitable area for identification
Solution Approach 1:
The patent applies preliminary action by having the learning model perform identification first, then using these results to systematically correct the teaching data. This preliminary identification step guides the subsequent correction process, ensuring that the final teaching data accurately indicates object areas while maintaining a relatively simple overall process.
Solution Approach 2:
The system uses feedback from the learning model's identification results to systematically improve teaching data quality. The identification accuracy metrics feed back into the correction process, automatically adjusting the teaching data to better indicate object areas without requiring complex manual intervention.
Data Source
AI summary
A teaching data correction device sets, for teaching data indicating an object area where an object of interest exists in a training image, a correction candidate area which is an area to be a correction candidate of the object area, the training image being used for learning. The teaching data correction device generates an output machine based on the correction candidate area, the output machine being learned to output, when an image is inputted thereto, an identification result or a regression result relating to the object of the interest. Then, the teaching data correction device updates the teaching data by the correction candidate area based on an accuracy of the output machine, the accuracy being calculated based on the identification result or the regression result outputted by the output machine.