Neural Network Training Using Trend-Guided Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning methods face limitations when the quantity of training data is small, leading to restricted accuracy in data augmentation, particularly for regression tasks where output values are sensitive to input changes.
Innovation Solution
A neural network training method that utilizes trend information to generate random training data by modifying the input range and masking output data, calculating a target loss based on the difference between prediction values to enhance training data diversity and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation technique is used to compensate for small training data quantity, then the training data quantity is increased, but the accuracy of data augmentation is limited due to regression issues where output values sensitively change according to input values
Solution Approach 1:
The patent applies parameter changes by modifying the input data range through trend information. Instead of randomly augmenting data within fixed ranges, the system adjusts input parameters based on learned trends from the neural network, allowing data augmentation to respect the underlying relationships in the data and reduce sensitive output changes
Solution Approach 2:
The patent implements feedback by using the neural network's prediction outputs and loss values to guide further data augmentation. The trend information derived from prediction differences is fed back into the data generation process, creating an iterative refinement loop that improves augmentation accuracy while increasing data quantity
2Adaptability or versatility
If conventional data augmentation methods modify input values within a fixed range, then data diversity is increased, but regression accuracy deteriorates because output values sensitively change with input changes
Solution Approach 1:
The patent applies dynamics by making the data augmentation range adaptive rather than static. The input modification range is dynamically adjusted based on trend information extracted from neural network predictions, allowing the system to explore more diverse input spaces in regions where the model is robust while being more conservative in sensitive regions
Solution Approach 2:
The system changes the parameters of data augmentation by using trend-based adjustments to input ranges. Instead of fixed or random modifications, the augmentation parameters are adapted based on the learned trends, maintaining data diversity while preserving regression accuracy
3Adaptability or versatility
If the input data range is expanded to improve data diversity, then data augmentation effectiveness is improved, but the reliability of prediction values deteriorates due to extrapolation beyond trained data distribution
Solution Approach 1:
The patent uses feedback from the neural network's own predictions to guide safe expansion of the input range. The trend information from prediction differences provides feedback on where expansion is likely to be reliable, allowing the system to improve data augmentation effectiveness while maintaining prediction reliability through informed extrapolation
Data Source
AI summary
A neural network training method and apparatus are disclosed, where the neural network training method using trend includes receiving original training data and trend information, obtaining a first prediction value and a first loss from a neural network based on the original training data, obtaining a second prediction value from the neural network based on the original training data using the trend information, determining a target loss by modifying the first loss based on the first prediction value and the second prediction value, and training the neural network based on the target loss.