Abstract:In this paper, we propose the Soft Generative Adversarial Network (SoftGAN), a strategy that utilizes a dynamic borderline softening mechanism to train Generative Adversarial Networks. This mechanism aims to solve the mode collapse problem and enhance the training stability of the generated outputs. Within the SoftGAN, the objective of the discriminator is to learn a fuzzy concept of real data with a soft borderline between real and generated data. This objective is achieved by balancing the principles of maximum concept coverage and maximum expected entropy of fuzzy concepts. During the early training stage of the SoftGAN, the principle of maximum expected entropy of fuzzy concepts guides the learning process due to the significant divergence between the generated and real data. However, in the final stage of training, the principle of maximum concept coverage dominates as the divergence between the two distributions decreases. The dynamic borderline softening mechanism of the SoftGAN can be likened to a student (the generator) striving to create realistic images, with the tutor (the discriminator) dynamically guiding the student towards the right direction and motivating effective learning. The tutor gives appropriate encouragement or requirements according to abilities of the student at different stages, so as to promote the student to improve themselves better. Our approach offers both theoretical and practical benefits for improving GAN training. We empirically demonstrate the superiority of our SoftGAN approach in addressing mode collapse issues and generating high-quality outputs compared to existing approaches.

Closed-Loop Training for Projected GAN

Understanding and Stabilizing GANs' Training Dynamics with Control Theory

Direct Adversarial Training for GANs

Direct Adversarial Training: A New Approach for Stabilizing The Training Process of GANs

Training Generative Adversarial Networks in One Stage

Incremental Focal Loss GANs.

Direct Adversarial Training: An Adaptive Method to Penalize Lipschitz Continuity of the Discriminator

A New Perspective on Stabilizing GANs Training: Direct Adversarial Training

Soft Generative Adversarial Network: Combating Mode Collapse in Generative Adversarial Network Training Via Dynamic Borderline Softening Mechanism

Improving GAN Training with Probability Ratio Clipping and Sample Reweighting

Simple Yet Effective Way for Improving the Performance of GAN

Class Balancing GAN with a Classifier in the Loop

Projected GANs Converge Faster

Stabilizing Training of Generative Adversarial Nets via Langevin Stein Variational Gradient Descent

CDL-GAN: Contrastive Distance Learning Generative Adversarial Network for Image Generation

CHAIN: Enhancing Generalization in Data-Efficient GANs via lipsCHitz continuity constrAIned Normalization

CL-GAN: A GAN-based Continual Learning Model for Generating and Detecting AGDs

Freeze the Discriminator: a Simple Baseline for Fine-Tuning GANs

Overcoming Catastrophic Forgetting for Fine-Tuning Pre-trained GANs

Fictitious GAN: Training GANs with Historical Models