基于生成对抗网络的变分自编码器解耦合

Decoupling of Variational Autoencoder Based on Generative Adversarial Network

深度生成模型从观测数据中学习到潜在因素,然后通过潜在因素生成目标,在人工智能领域受到广泛关注。现有深度生成模型学习的潜在因素往往是耦合的,无法让潜在因素每一维控制所得数据的不同特征,即无法单独改变某一特征而不影响其他特征。为此,在β-变分自编码器(beta-variational autoencoder,β-VAE)的基础上,结合生成对抗网络(generative adversarial networks,GAN),提出基于生成对抗网络的变分自编码器(beta-variational autoencoder based on generative adversarial network,β-GVAE)模型。该模型是一种改进的β-VAE,通过引入生成对抗网络约束β-VAE中损失函数的KL项(Kullback-Leibler divergence),促进模型的解耦合。在数据集CelebA、3D Chairs和dSprites上进行对比实验,结果表明β-GVAE不仅具有更好的解耦合表示,同时生成的图像具有更好的视觉效果。

Deep generative models learn latent factors from observational data,and then generate targets through latent factors,which have received extensive attention in the field of artificial intelligence.The latent factors learned by the existing deep generative models are often coupled,and each dimension of the latent factors cannot control different characteristics of the obtained data,that is,it is impossible to change a certain characteristic independently without affecting other characteristics.Therefore,beta-variational autoencoder(β-VAE)based on generative adversarial network(β-GVAE)is proposed based on β-VAE and combined with generative adversarial networks(GAN).This model is an improved β-VAE,which promotes the decoupling of the model by introducing a generative adversarial network to constrain the KL divergence of the loss function in β-VAE.By designing comparative experiments on three datasets,CelebA,3D Chairs and dSprites,it is proved thatβ-GVAE not only has better decoupled representation,but also the generated images has better visual effects.