基于监督对比正则化项的信息蒸馏留生成对抗网络  

Information Distillation Generative Adversarial Net Based on Supervised Contrastive Regularization

作　　者：陈亚瑞[1] 王晓捷 李晴[1] 刘浩天 史艳翠 赵婷婷[1] CHEN Yarui;WANG Xiaojie;LI Qing;LIU Haotian;SHI Yancui;ZHAO Tingting(College of Artificial Intelligence,Tianjin University of Science and Technology,Tianjin 300457,China)

机构地区：[1]天津科技大学人工智能学院,天津300457

出　　处：《天津科技大学学报》2025年第2期61-70,共10页Journal of Tianjin University of Science & Technology

基　　金：国家自然科学基金项目(61976156)。

摘　　要：传统生成对抗网络主要通过最大化解耦表示和生成数据之间的互信息来学习解耦表示,较少分析解耦表示各维度之间的独立性。本文提出一种基于监督对比正则化项的信息蒸馏生成对抗网络(information distillation generative adversarial net based on supervised contrastive regularization,IDGAN-SC)。首先,IDGAN-SC模型利用β-VAE模型学习解耦表示空间,约束解耦表示空间和生成模型之间具有强相关性;然后,通过最大化解耦隐向量和生成数据之间的互信息对模型进行解耦表示学习,进一步利用监督对比正则化项的对比分类信息增强解耦隐变量各维度之间的独立性。在dSprites、MNIST、CelebA数据集上,分别从定性和定量的角度设计了对比实验,实验结果表明相比已有的生成对抗网络的解耦性能,IDGAN-SC模型具有较强的解耦能力并具有明显的解耦效果。For generative adversarial net(GAN),traditional methods mainly maximize disentangling the latent presentation based on the mutual information between the disentangled representation and the generated data,but rarely analyze the independence among dimensions of the latent vector.In this article,we propose an information distillation generative adversarial net based on supervised contrastive regularization(IDGAN-SC).The IDGAN-SC model firstly learns disentangled representation space through training β-VAE,which enforces strong correlation between the disentangled representation space and the generative model.Then,the model constructs the disentangled structure by maximizing the mutual information between the disentangled latent vectors and the generated data.Furthermore,the model utilizes the contrastive classification information of the supervised contrastive regularization to enhance the independence between dimensions of the latent vectors.In our present study,we conducted quantitative and qualitative experiments on the dSprites,MNIST,and CelebA datasets.Experi-ments showed that IDGAN-SC significantly outperformed current disentanglement methods based on the disentanglement metrics.

关 键 词：生成对抗网络 变分自编码器 解耦表示 对比学习 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]