Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Fast-Slow Co-advancing Optimizer: Toward Harmonious Adversarial Training of GAN

Lin Wang, Xiancheng Wang, Rui Wang, Zhibo Zhang, Minghang Zhao

TL;DR

GAN training remains unstable due to hyperparameter sensitivity in the dual-player setup. FSCO integrates Deep Deterministic Policy Gradient (DDPG) as an embedded agent to adapt the discriminator's step size in real time, using the relation $\eta_{FSCO-D}(t) = \eta_{D}(t) \times u(t)$ and a reward $Reward(t) = -|Gloss(t) - Dloss(t)|$ to align generator and discriminator learning. The method smooths the adversarial dynamics, expanding the stable hyperparameter region and reducing the need for manual tuning, as demonstrated on MNIST, ANIME, and Ganyu. However, some limitations persist, including occasional overfitting and failures at higher resolutions, indicating areas for refinement and scope for diffusion-model alternatives.

Abstract

Up to now, the training processes of typical Generative Adversarial Networks (GANs) are still particularly sensitive to data properties and hyperparameters, which may lead to severe oscillations, difficulties in convergence, or even failures to converge, especially when the overall variances of the training sets are large. These phenomena are often attributed to the training characteristics of such networks. Aiming at the problem, this paper develops a new intelligent optimizer, Fast-Slow Co-advancing Optimizer (FSCO), which employs reinforcement learning in the training process of GANs to make training easier. Specifically, this paper allows the training step size to be controlled by an agent to improve training stability, and makes the training process more intelligent with variable learning rates, making GANs less sensitive to step size. Experiments have been conducted on three benchmark datasets to verify the effectiveness of the developed FSCO.

Fast-Slow Co-advancing Optimizer: Toward Harmonious Adversarial Training of GAN

TL;DR

GAN training remains unstable due to hyperparameter sensitivity in the dual-player setup. FSCO integrates Deep Deterministic Policy Gradient (DDPG) as an embedded agent to adapt the discriminator's step size in real time, using the relation and a reward to align generator and discriminator learning. The method smooths the adversarial dynamics, expanding the stable hyperparameter region and reducing the need for manual tuning, as demonstrated on MNIST, ANIME, and Ganyu. However, some limitations persist, including occasional overfitting and failures at higher resolutions, indicating areas for refinement and scope for diffusion-model alternatives.

Abstract

Up to now, the training processes of typical Generative Adversarial Networks (GANs) are still particularly sensitive to data properties and hyperparameters, which may lead to severe oscillations, difficulties in convergence, or even failures to converge, especially when the overall variances of the training sets are large. These phenomena are often attributed to the training characteristics of such networks. Aiming at the problem, this paper develops a new intelligent optimizer, Fast-Slow Co-advancing Optimizer (FSCO), which employs reinforcement learning in the training process of GANs to make training easier. Specifically, this paper allows the training step size to be controlled by an agent to improve training stability, and makes the training process more intelligent with variable learning rates, making GANs less sensitive to step size. Experiments have been conducted on three benchmark datasets to verify the effectiveness of the developed FSCO.

Paper Structure

This paper contains 23 sections, 7 equations, 11 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Generative Adversarial Networks (GAN)
  4. Reinforcement Learning(RL)
  5. Algorithm Introduction
  6. Basics of GAN
  7. Basics of DDPG
  8. The developed FSCO
  9. Automatic Setting of Discriminator Training Step Size
  10. DDPG Reward Setting
  11. Overall Steps
  12. Experiments
  13. Experiment on 28×28 Image Dataset
  14. Hyperparameter Setup
  15. Experimental Results
  16. ...and 8 more sections

Figures (11)

  • Figure 1: Dialectical thinking guided diagram
  • Figure 2: The Structure of GAN
  • Figure 3: The Structure of DDPG
  • Figure 4: Harmonious adversarial training diagram in terms of step size
  • Figure 5: Algorithm diagram of the developed FSCO method
  • ...and 6 more figures