Adversarial Learning Device Frequency Domain Discrimination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep generation models, such as GANs, suffer from over-learning due to the inclusion of high frequency components in generated samples, leading to reduced accuracy and deteriorated generation quality.
Innovation Solution
A learning device that includes a conversion unit for converting data into frequency components, a calculation unit for optimizing the generator and discriminators using a loss function that balances spatial and frequency domain performance, and an update unit for updating model parameters to suppress over-learning and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a conventional GAN model is used for data generation, then the model can generate samples, but high frequency components not included in actual learning data are mixed in the generated samples causing over-learning
Solution Approach 1:
The discriminator is divided into two separate discriminators: one for the spatial domain and one for the frequency domain. This segmentation allows each discriminator to focus on specific aspects of the generated data, with the frequency domain discriminator specifically targeting high frequency components to prevent over-learning while the spatial domain discriminator handles overall structure.
Solution Approach 2:
The patent introduces a frequency domain dimension by applying Fourier transform to the generated images. Instead of only evaluating samples in the spatial domain, the system transforms samples into the frequency domain and evaluates them separately, adding a new dimension of analysis that captures high frequency information otherwise invisible in the spatial domain.
2Measurement precision
If the discriminator focuses on high frequency components, then over-learning is detected, but the model accuracy deteriorates due to over-learning
Solution Approach 1:
The frequency domain discriminator acts as an intermediary that specifically evaluates high frequency components before the main training process. By introducing this intermediate evaluation step, the system can detect over-learning tendencies in high frequency regions and adjust training accordingly, preventing the degradation of overall model accuracy.
Solution Approach 2:
The patent changes the evaluation parameters by transforming images from spatial domain to frequency domain using Fourier transform. This parameter change allows the discriminator to operate on different features (frequency components rather than pixel values), enabling detection of over-learning without directly compromising the spatial domain generation quality.
3Device complexity
If only spatial domain discrimination is used, then the training process is simple, but the discriminator cannot detect over-learning in high frequency components
Solution Approach 1:
The patent merges two discrimination processes: spatial domain discrimination and frequency domain discrimination. Both discriminators work together in a unified training framework where the frequency domain discriminator specifically targets high frequency over-learning while the spatial domain discriminator handles general structure, combining their strengths for comprehensive evaluation.
Solution Approach 2:
The frequency domain analysis is integrated continuously into the training process rather than being a separate post-processing step. The Fourier transform and frequency domain discrimination occur continuously during training, providing ongoing detection of over-learning in high frequency components throughout the entire training trajectory.
Data Source
AI summary
The conversion unit (132) converts first data into a first frequency component, and converts second data generated by a generator that configures an adversarial learning model into a second frequency component. The calculation unit (133) calculates a loss function that simultaneously optimizes the generator, a first discriminator that configures the adversarial learning model and discriminates between the first data and the second data, and a second discriminator that configures the adversarial learning model and discriminates between the first frequency component and the second frequency component. The update unit (134) updates parameters of the generator, the first discriminator, and the second discriminator so that the loss function calculated by the calculation unit (133) is optimized.