Stabilizing GAN Discriminator via Gradient-Based False Data Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative adversarial networks face unstable learning, where the generator may produce false data that significantly differs from previous data, leading to penalization of the discriminator and unstable model generation.
Innovation Solution
A model generation system that includes a saved data storage unit, a data generation unit, a discriminator unit, a gradient information calculation unit, and a selection unit, which calculates distances between true and false data output values and updates weights to increase discrimination accuracy, selecting and storing false data based on gradient information to stabilize model generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the generator generates false data randomly, then the generator can produce diverse data, but the discriminator becomes unstable and learns to discriminate specific false data with high accuracy
Solution Approach 1:
The patent applies preliminary action by calculating gradient information for all generated false data before the discriminator updates its model. This pre-calculation of gradients ensures that even if the discriminator temporarily achieves high accuracy on certain false data, the model updates are based on aggregated gradient information from all data points, preventing unstable learning from occurring
Solution Approach 2:
The patent implements feedback by using the calculated gradient information to guide both the discriminator and generator updates. The gradient information provides feedback on how well the discriminator is performing across all false data, allowing the system to adjust and maintain stable learning while preserving data diversity
2Measurement precision
If the discriminator discriminates false data with very strong false data-like characteristics, then discrimination accuracy improves, but the model learns to discriminate only specific false data, causing unstable learning
Solution Approach 1:
The patent uses gradient information as feedback to prevent the discriminator from over-fitting to specific false data patterns. By calculating gradients across all false data and using this aggregated information for model updates, the system maintains discrimination accuracy while ensuring stable, generalizable learning
Solution Approach 2:
The patent changes the parameter update mechanism by using gradient information to determine weight updates rather than simple binary discrimination outcomes. This continuous gradient-based parameter adjustment allows the discriminator to maintain high accuracy while avoiding unstable learning patterns
Data Source
AI summary
A data generation unit 6 generates a first number of false data based on a generation model that is a neural network for generating false data. A discriminator unit 7 derives output values for given data based on a discriminator model that is a neural network for deriving output values indicating true data-likeness and false data-likeness of the given data. A gradient information calculation unit 8 calculates, for each combination of one true data and each of the first number of false data, a distance between the output value for the true data and the output value for the false data, and calculates gradient information that is an update amount for each weight that the discriminator model has, so as to increase the distance by a predetermined amount. A selection unit 9 selects false data to be stored in the saved data storage unit 4.