Abstract:Generative adversarial networks (GANs) have emerged as a powerful tool for generating high-fidelity data. However, the main bottleneck of existing approaches is the lack of supervision on the generator training, which often results in undamped oscillation and unsatisfactory performance. To address this issue, we propose an algorithm called Monte Carlo GAN (MCGAN). This approach, utilizing an innovative generative loss function, termly the regression loss, reformulates the generator training as a regression task and enables the generator training by minimizing the mean squared error between the discriminator's output of real data and the expected discriminator of fake data. We demonstrate the desirable analytic properties of the regression loss, including discriminability and optimality, and show that our method requires a weaker condition on the discriminator for effective generator training. These properties justify the strength of this approach to improve the training stability while retaining the optimality of GAN by leveraging strong supervision of the regression loss. Numerical results on CIFAR-10 and CIFAR-100 datasets demonstrate that the proposed MCGAN significantly and consistently improves the existing state-of-the-art GAN models in terms of quality, accuracy, training stability, and learned latent space. Furthermore, the proposed algorithm exhibits great flexibility for integrating with a variety of backbone models to generate spatial images, temporal time-series, and spatio-temporal video data.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve the main bottleneck problem in the training of Generative Adversarial Networks (GANs), that is, the lack of supervision in generator training, which usually leads to undamped oscillations and unsatisfactory performance. Specifically: 1. **Insufficient supervision in generator training**: Existing GAN methods lack effective supervision during the generator training process, which may lead to instability and poor performance during training. 2. **Oscillation problems during training**: The adversarial training between the generator and the discriminator is prone to cause undamped oscillations, making it difficult for the model to converge. 3. **Unstable generation quality**: The quality of data generated by the generator may be inconsistent, especially in high - fidelity data generation tasks. To address these problems, the authors propose a new algorithm - Monte Carlo GAN (MCGAN). MCGAN re - formulates the generator training as a regression task by introducing an innovative generation loss function (called regression loss), thereby enhancing the training stability of the generator and improving the generation quality. ### Main contributions 1. **Proposing the MCGAN method**: This method enhances the training of unconditional and conditional GANs by introducing regression loss. 2. **Establishing the theoretical basis of regression loss**: The authors prove the discriminability and optimality of the regression loss and demonstrate the improvement of MCGAN in training stability. 3. **Experimentally proving the effectiveness of MCGAN**: Through experiments on the CIFAR - 10 and CIFAR - 100 datasets, the consistency and superiority of MCGAN in image generation, time - series generation and video generation tasks are proved. ### Key techniques - **Regression loss**: The regression loss \(L_R\) optimizes the generator by minimizing the mean - square error between the discriminator's output for real data and the expected fake data: \[ L_R(\theta; \phi):=\mathbb{E}_{(X, Y)\sim\mu}\left[|D_{\phi}(X)-\mathbb{E}_{x\sim\nu_{\theta}(Y)}[D_{\phi}(x)]|^2\right] \] - **Monte Carlo estimation of discriminator output**: Use the Monte Carlo method to estimate the discriminator output under the fake data distribution, thus avoiding complex conditional expectation calculations. - **Training stability**: Through strongly - supervised regression loss, MCGAN can train the generator more efficiently and stably under weaker discriminator conditions. ### Experimental results - **Image generation**: On the CIFAR - 10 and CIFAR - 100 datasets, MCGAN significantly improves the quality, accuracy and training stability of generated images. - **Time - series generation**: MCGAN shows good flexibility and performance in time - series data generation tasks. - **Video generation**: MCGAN also shows superior performance in video data generation tasks, especially in generating high - quality and diverse video data. ### Conclusion MCGAN effectively solves the problems of insufficient supervision and training instability in GAN training by introducing regression loss, and significantly improves the performance of the generator. This method not only performs well in image generation tasks, but also is suitable for various tasks such as time - series and video data generation.

MCGAN: Enhancing GAN Training with Regression-Based Generator Loss

RAGAN: Regression Attention Generative Adversarial Networks

Improving MMD-GAN Training with Repulsive Loss Function

Improving GAN Training via Feature Space Shrinkage

Improving GAN with Neighbors Embedding and Gradient Matching

Generative Adversarial Networks with Learnable Auxiliary Module for Image Synthesis

Training Generative Adversarial Networks by Auxiliary Adversarial Example Regulator

Teacher-Supervised Generative Adversarial Networks.

Towards the Gradient Vanishing, Divergence Mismatching and Mode Collapse of Generative Adversarial Nets

Addressing GAN Training Instabilities via Tunable Classification Losses

Evaluation of GAN-Based Model for Adversarial Training

Enhancing Stability in Training Conditional Generative Adversarial Networks via Selective Data Matching

Generative Adversarial Networks with Denoising Penalty and Sample Augmentation

Improvement and Application of Generative Adversarial Networks Algorithm Based on Transfer Learning

On the Effectiveness of Least Squares Generative Adversarial Networks

Optimizing Generative Adversarial Networks in Latent Space

Gang of GANs: Generative Adversarial Networks with Maximum Margin Ranking

The residual generator: An improved divergence minimization framework for GAN

Generalizing Semi-Supervised Generative Adversarial Networks to Regression Using Feature Contrasting

An Empirical Study on GANs with Margin Cosine Loss and Relativistic Discriminator

Simple Yet Effective Way for Improving the Performance of GAN