Combined GAN for Synthetic Data Fabrication
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing generative adversarial networks (GANs) struggle to generate synthetic data that both imitates the characteristics of original structured data and adheres to user-defined constraints, especially when the original data does not fully meet the desired requirements.
Innovation Solution
A combined GAN is trained using both original structured data and fabricated data that adhere to user-defined constraints, comprising a generator and two static discriminators, where the generator produces data examples that are evaluated by both discriminators, and the logic AND gate provides feedback until the generated data is deemed plausible, ensuring the output meets both data characteristics and constraints.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single GAN is trained on original structured data, then the generated data imitates the characteristics of original data, but the generated data cannot adhere to user-defined constraints
Solution Approach 1:
The patent divides the GAN system into two independent parts: a first GAN trained on original structured data to capture data characteristics, and a second GAN trained on fabricated data adhering to user-defined constraints. This segmentation allows each GAN to specialize in one function, resolving the contradiction between characteristic imitation and constraint adherence.
Solution Approach 2:
The patent merges the two separate GANs into a combined GAN system where the generator from the first GAN is combined with the discriminator from the second GAN. This merging enables the system to simultaneously achieve both data characteristic imitation (from the first GAN's generator) and constraint adherence (from the second GAN's discriminator).
2Adaptability or versatility
If a combined GAN is trained using both original and fabricated data, then the generated data adheres to user-defined constraints, but the system complexity increases
Solution Approach 1:
The combined GAN system is designed to perform multiple functions: the generator imitates data characteristics while the discriminator ensures constraint adherence. By making the system multi-functional, the patent reduces the need for separate systems, thereby managing complexity while achieving versatility.
Solution Approach 2:
The patent introduces an intermediary mechanism where the generator's output is evaluated by both discriminators, and the results are combined using a logic AND gate. This intermediary evaluation mechanism simplifies the training process by providing a clear combined feedback signal, making the complex system more manageable.
3Manufacturing precision
If the generator is trained to satisfy both data characteristics and constraints simultaneously, then the generated data meets all requirements, but the training process becomes more difficult
Solution Approach 1:
The patent performs preliminary training of two separate GANs before combining them. The first GAN is pre-trained on original data to learn characteristics, and the second GAN is pre-trained on constrained data to learn constraint satisfaction. This preliminary action simplifies the final combined training by pre-establishing the necessary learning foundations.
Solution Approach 2:
The combined GAN uses a feedback mechanism where the logic AND gate combines the evaluations from both discriminators and provides unified feedback to the generator. This feedback approach simplifies training by consolidating multiple constraint signals into a single clear guidance signal for the generator.
Data Source
AI summary
A computer-implemented method including: training a first Generative Adversarial Network (GAN) based on original structured data; training a second GAN based on fabricated structured data that adhere to user-defined constraints; combining the first and second GANs into a combined GAN; training the combined GAN; and operating the trained combined GAN to generate new fabricated data that both imitate characteristics of the original structured data, and adhere to the user-defined constraints.