基于Wasserstein距离的双向学习推理

基于Wasserstein距离的生成对抗网络(WGAN)将编码器和生成器双向集成于其模型中,从而增强了生成模型的学习能力,但其在优化目标中使用KL散度度量分布间的差异,会导致学习训练过程中出现梯度消失或梯度爆炸问题,降低模型鲁棒性.为克服这一问题,提出了一种基于Wasserstein距离的双向学习推理(WBLI)模型.文章首先建立了真实数据分布与隐数据分布双向学习网络,然后引入Wasserstein距离度量联合概率分布的差异性,并据此推导了可解的损失代价函数,给出了完整的网络学习模型和迭代算法.实验结果表明,WBLI模型有效缓解了传统GAN及其变种的模式坍塌问题,增强了训练学习的鲁棒性,可生产辨识度更高的样本.

Bidirectional learned inference based on Wasserstein distance

In WGAN, embedding encoder into Generative Adversarial Networks(GAN)can enhance the learning ability of the generative model. However, using the Kullback-Leibler(KL)divergence to measure the difference between two distributions in the optimization objective will lead to the gradient vanishing or gradient explosion problem in the learning training process and reduce the robustness of model. In order to tackle this problem, a Wasserstein-distance-based Bidirectional Learned Inference(WBLI)model is proposed in this paper. A bidirectional network is first established for learning the distribution of the true data and latent variables, where the difference of the joint probability distribution is measured by the Wasserstein distance. Based on this Wasserstein distance, we redesign the loss function which is solvable and consequently propose an iterative algorithm. The experimental results show that the WBLI model overcomes the defects of traditional GAN and its variants. It effectively eliminates the model collapse problem of generating models, increases the robustness of training learning, and contributes to the improvement of the recognition rate of classifiers.