Generative Adversarial Network Training Using Deterministic Feature Assessment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for training generative adversarial networks to evaluate expert knowledge are less stable and require additional resources, leading to longer training times and increased resource consumption.
Innovation Solution
The method involves using deterministic functions instead of neural networks for evaluating specific features, allowing for faster and more robust training by applying these functions to assess the realism of generated data and reducing the need for separate training, thereby stabilizing and reproducibility of the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If an additional neural network is provided to verify specific features in the discriminator, then the discriminator can evaluate expert knowledge and specific features, but the training process becomes less stable and requires additional resources and time
Solution Approach 1:
The patent extracts the specific feature verification function from a separate neural network and integrates it into the existing discriminator. The discriminator now directly evaluates both general realism and specific features (such as physics consistency, expert knowledge compliance) within a single network architecture, eliminating the need for additional neural networks while maintaining evaluation accuracy
Solution Approach 2:
The patent merges multiple evaluation functions (realism assessment and specific feature verification) into a single discriminator network. The discriminator simultaneously performs both tasks using unified architecture and loss functions, combining what were previously separate evaluation mechanisms into one integrated system
2Reliability
If an additional neural network is provided to verify specific features, then feature evaluation capability is enhanced, but training time and resource consumption increase
Solution Approach 1:
The patent removes the need for separate training processes for additional verification networks by integrating feature evaluation directly into the discriminator. This eliminates redundant training iterations and reduces overall training time while maintaining comprehensive feature verification capability
Solution Approach 2:
The discriminator serves multiple functions simultaneously - it both generates realism assessments and verifies specific features without requiring external verification networks. This self-sufficient approach reduces computational overhead and accelerates training by eliminating dependency on additional trained components
3Reliability
If an additional neural network is provided for specific feature verification, then evaluation comprehensiveness is improved, but reproducibility of training decreases
Solution Approach 1:
The patent extracts the source of variability by removing additional trainable neural networks from the architecture. By evaluating specific features directly within the discriminator using deterministic criteria rather than separate learned models, the system achieves more consistent and reproducible training results across different runs
Data Source
AI summary
A method for training a generative adversarial network. The method includes: iteratively training the generative adversarial network based on the training data, the training of the generative adversarial network including an alternating training of the generator and the discriminator based on the training data, the training of the generator including a training the generator based on the training data and results of realism assessments performed by the discriminator, and the training of the generative adversarial network in each iteration step including a generation of corresponding data by the generator, a performance of a realism assessment of the corresponding data by the discriminator, and a performance of an additional realism assessment of at least one specific feature derived from the corresponding data by the discriminator, and the performance of the additional realism assessment of at least one specific feature derived from the corresponding data including an application of a deterministic function.