生成对抗网络(Generative Adversarial Network,GAN)是目前深度学习领域的一个研究热点.针对GAN生成模型存在模式坍塌和训练不稳定的问题,提出一种全新的能量函数意义下的生成式对抗网络模型(En-ergy Maximum Entropy GAN,E-MEGAN).该模型的最终目标是最大化生成样本的熵来解决模式坍塌问题,使用非参数互信息估计量计算该熵.为了稳定对抗的训练过程,还使用了以零为中心的梯度惩罚技巧.通过在MNIST和CelebA数据集上进行大量实验,表明该模型可以生成清晰高质量的图像,其IS(Inception Scores)和FID(Frechet Inception Distance)与WGAN-GP技术相比具有同等竞争力,并且不会遭受模式的损失.
MAXIMUM ENTROPY GENERATIVE ADVERSARIAL NETWORK BASED ON ENERGY MODELS
Generative adversarial network(GAN)is a hot topic in the field of deep learning.Aiming at the problem of GAN model such as mode collapse and training instability in the training process,we propose a new energy maximum entropy GAN(E-MEGAN).The ultimate goal of the model was to maximize the entropy of the generated samples to solve the problem of mode collapse,and we used the non-parametric mutual information estimator to calculate the entropy.At the same time,in order to stabilize the adversarial training process,we also used a zero-centered gradient penalty technique.Through a large number of experiments on the MNIST and CelebA datasets,it is shown that this model can generate clear and high-quality images,and its inception scores(IS)and Fréchet inception distance(FID)are equally competitive compared with WGAN-GP technology.The method also will not suffer from mode collapse.
Generative adversarial networkEnergy functionMaximum entropyStable against
万方数据
维普
