Training Data Generation via Subgroup Discriminator Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of machine learning discriminators largely depends on the quantity and uniformity of training data, and manual annotation methods for generating correct answer data are time-consuming and prone to quality variations when performed by multiple workers.
Innovation Solution
A method involving the division of training image groups into subgroups, where first correct answer data is generated independently for each subgroup, and a discriminator is learned for each subgroup. The discriminator's output is used to generate second correct answer data by integrating with the first correct answer data, reducing quality variations through averaging or statistical processing across subgroups.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual annotation method is used to generate correct answer data, then the work can be completed by multiple workers or divided into multiple times, but the variation in the quality of the generated correct answer data will occur
Solution Approach 1:
The training image group is divided into multiple image subgroups, and correct answer data is generated independently for each subgroup. This segmentation allows multiple workers to work on different subgroups simultaneously, improving productivity while maintaining quality control through consistent processing methods across all subgroups
Solution Approach 2:
A discriminator is learned based on the first correct answer data from each image subgroup, and this discriminator is then used to generate second correct answer data that serves as feedback. This feedback mechanism helps standardize the quality of correct answer data across different workers and subgroups, reducing quality variation
2Reliability
If manual annotation method is used to generate correct answer data, then the correct answer data can be generated, but the work time becomes long
Solution Approach 1:
The training image group is pre-divided into multiple image subgroups before annotation begins. This preliminary action enables parallel processing of correct answer data generation across different subgroups by multiple workers simultaneously, significantly reducing the total annotation time while maintaining data accuracy
Solution Approach 2:
The first correct answer data from multiple workers and the second correct answer data generated by the discriminator are merged to create the final correct answer data. This combination approach maintains high accuracy by leveraging multiple sources while reducing time through parallel processing
Data Source
AI summary
A training data generation device performs the processes of: acquiring first correct answer data that is correct answer data with respect to each training image of a training image group, the training image group including plural training images to be given the correct answer data, the training image group being divided into plural image subgroups, the first correct answer data being generated in a different situation for each of the plural image subgroups; generating a discriminator with respect to the each of the plural image subgroups, the discriminator being learned based on the each of the plural image subgroups and the first correct answer data corresponding to the each of the plural image subgroups; acquiring a discrimination result outputted by the discriminator for each of the plural training images by inputting the each of the plural training images to the discriminator which is generated based on at least one image subgroup other than an image subgroup to which the each of the training images belongs; and generating second correct answer data for the each of the plural training images based on the discrimination result and the first correct answer data.