Related Concept Videos
Random Sampling Method
Sampling Continuous Time Signal
In the...
Randomized Experiments
Simple randomization
Simple...
Sampling Distribution
Residuals and Least-Squares Property
If the observed data point lies above the line, the residual is positive, and the line underestimates the actual data value for y. If the observed data point lies below the line, the residual is negative, and the line overestimates the actual data value for y.
The process of fitting the best-fit...
Sampling Theorem
You might also read
Related Articles
Articles linked to this work by shared authors, journal, and citation graph.
Serum concentrations of antioxidant vitamins and carotenoids are low in individuals with a history of attempted suicide.
Arsenic trioxide induces different gene expression profiles of genes related to growth and apoptosis in glioma cells dependent on the p53 status.
Role of p38 mitogen-activated protein kinases in cardioprotection of morphine preconditioning.
Related Experiment Video
Updated: Oct 18, 2025

Generating Strictly Controlled Stimuli for Figure Recognition Experiments
Published on: March 18, 2019
Real Sample Consistency Regularization for GANs.
1College of Sciences, Northeastern University, Shenyang 110819, China.
Real Sample Consistency (RSC) regularization addresses mode collapse in generative adversarial networks by preventing discriminator misjudgment. This method stabilizes training and improves image generation quality, outperforming existing techniques.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Mode collapse is a persistent challenge in generative adversarial networks (GANs).
- Zero Gradient Penalty (0GP) regularization mitigates mode collapse but can worsen discriminator misjudgment, where generated samples are incorrectly perceived as more real than actual data.
- This discriminator misjudgment leads to unnatural image generation and reduced quality.
Purpose of the Study:
- To introduce Real Sample Consistency (RSC) regularization as a novel method to address discriminator misjudgment in GANs.
- To improve the stability and quality of GAN training and image generation.
- To provide a more effective alternative to existing regularization techniques like 0GP.
Main Methods:
- Proposed Real Sample Consistency (RSC) regularization.
- RSC involves randomly dividing real samples into two groups during training.
- The method minimizes the loss between the discriminator's outputs for these two groups, enforcing consistent outputs for all real samples.
Main Results:
- RSC effectively alleviates discriminator misjudgment, leading to more stable GAN training compared to 0GP regularization.
- Significant improvements in Frechet Inception Distance (FID) scores were observed: from 14.28 to 9.8 on CIFAR-10 (FARGAN), 23.42 to 17.14 on CIFAR-100, and 53.79 to 46.92 on ImageNet2012.
- The average distance between generated and real samples decreased from 0.028 to 0.025 on synthetic data.
- Generator and discriminator losses in standard GANs with RSC approached theoretical values and remained stable.
Conclusions:
- Real Sample Consistency (RSC) regularization is a highly effective method for improving GAN performance by tackling discriminator misjudgment.
- RSC offers a more stable training process and superior generation quality compared to 0GP regularization.
- The proposed method demonstrates broad applicability and significant performance gains across various datasets and GAN architectures.