基于互信息最大化正则的生成式对抗网络

Generative Adversarial Nets Based on Mutual Information Maximization Regularity

目前生成式对抗网络(GAN)面临的一大难点是模型训练过程中普遍存在着不收敛性,其最为突出的表现即"模式坍塌"现象,从而导致生成的目标对象缺乏多样性。首先从信息论角度简要解释原始GAN的理论方法,在此基础之上,提出一种基于互信息最大化的正则方法,通过最大化随机噪声变量和生成样本之间的互信息,为生成网络的目标函数提供一个上界函数,迫使生成的数据分布更加趋于真实数据分布。对MNIST手写字符数据进行的实验结果表明,引入互信息正则,有助于提升生成样本的多样性。

Currently,one of the great difficulty of Generation adversarial nets（GAN）is the prevalence of non-convergence in the training dynamics of model,and its most prominent manifestation is the mode collapse phenomenon,which leads to the lack of diversity of the target object.First explains the theoretical method of the original GAN from the perspective of information theory.On this basis,proposes a regular method based on maximization of the mutual information.By maximizing the mutual information between the noise source and the generated samples,it provides an upper bound function of the objective that forces the produced data distribution to converge more toward the real data distribution.The experimental results of MNIST data show that the maximization of the mutual information introduced is useful to increase the diversity of samples.