Learning Data Generation Using Neural Array Matching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Creating teaching data for deep learning models, such as lesion area recognition in endoscopic images, requires significant manual annotation effort, which is time-consuming and labor-intensive, and existing data augmentation methods fail to account for real-world environmental changes.
Innovation Solution
A learning data generating apparatus and method that utilizes manually annotated training data to automatically generate correct answer information for unannotated data by comparing vector outputs from a neural network, reducing the need for extensive manual annotation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual annotation is used to create teaching data, then the accuracy of correct answer information is improved, but the time and labor required increases significantly
Solution Approach 1:
The patent applies preliminary action by using manually annotated data to train a neural network in advance. This trained network then automatically generates correct answer information for unannotated data, eliminating the need for manual annotation of each individual dataset while maintaining accuracy through the pre-trained model's predictive capabilities
Solution Approach 2:
The patent uses copying by replicating the patterns and features learned from manually annotated training data through the neural network. The network copies the essential characteristics of correctly annotated regions and applies them to generate correct answer information for new, unannotated images, thereby scaling accuracy without proportional increases in manual labor
2Measurement precision
If more training data is collected to improve model performance, then the accuracy of the neural network is improved, but the amount of manual annotation work increases
Solution Approach 1:
The patent implements self-service by enabling the neural network to generate its own training data with automatic correct answer information. The system uses the trained network to process unannotated images and generate accurate correct answer information automatically, allowing the system to expand its training dataset without requiring proportional increases in manual annotation resources
Solution Approach 2:
The patent applies parameter changes by transitioning from manual annotation parameters to automated neural network prediction parameters. This involves changing the method of generating correct answer information from human-driven manual marking to algorithm-driven automatic generation, thereby improving productivity while maintaining or enhancing accuracy through the network's learned parameters
Data Source
AI summary
A learning data generating apparatus receives first training data, first correct answer information, and second training data, inputs the first and second training data into a neural network to thereby cause the neural network to output first and second array groups each constituted of a plurality of arrays, identifies a first array from the first array group based on the first correct answer information, the first array being an array corresponding to the recognition target, and compares the first array and each of the plurality of arrays constituting the second array group, to create second correct answer information corresponding to the second training data.