人工智能培训

搜索

人工智能论文:多样性敏感的条件生成对抗网络(Diversity-Sensitive Conditional Generative Adversarial Net

[复制链接]
mrlam 发表于 2019-1-28 11:10:59 | 显示全部楼层 |阅读模式
mrlam 2019-1-28 11:10:59 112 0 显示全部楼层
人工智能论文:多样性敏感的条件生成对抗网络(Diversity-Sensitive Conditional Generative Adversarial Networks)我们提出了一种简单但高效的方法来解决条件生成对抗网络(cGAN)中的模糊崩溃问题。虽然条件分布是多模态的(即,有很多模式)实践,但大多数cGAN方法倾向于学习过于简化的分布。无论代码的变化如何,输入始终映射到单个输出。为了解决这个问题,我们建议明确规范发电机根据潜码产生不同的输出。所提出的规范化简单,通用,并且可以容易地集成到大多数有条件的GAN目标中。此外,关于生成器的明确正规化我们控制视觉质量和多样性之间平衡的方法。我们展示了我们的方法在三个条件生成任务上的有效性:图像到图像的翻译,图像修复和未来的视频预测。我们表明,对现有模型简单地加上我们的正则化导致了令人惊讶的多样化世代,大大优于以前在每个单独任务中专门设计的多模态条件生成方法。
We propose a simple yet highly effective method that addresses themode-collapse problem in the Conditional Generative Adversarial Network (cGAN).Although conditional distributions are multi-modal (ie, having many modes) inpractice, most cGAN approaches tend to learn an overly simplified distributionwherean input is always mapped to a single output regardless of variations inlatent code.To address such issue, we propose to explicitly regularize thegenerator to produce diverse outputs depending on latent codes.The proposedregularization is simple, general, and can be easily integrated into mostconditional GAN objectives.Additionally, explicit regularization on generatorallows our method to control a balance between visual quality and diversity.Wedemonstrate the effectiveness of our method on three conditional generationtasks: image-to-image translation, image inpainting, and future videoprediction.We show that simple addition of our regularization to existingmodels leads to surprisingly diverse generations, substantially outperformingthe previous approaches for multi-modal conditional generation specificallydesigned in each individual task.人工智能论文:多样性敏感的条件生成对抗网络(Diversity-Sensitive Conditional Generative Adversarial Networks) vG99vh7j71d3qrvR.jpg
URL地址:https://arxiv.org/abs/1901.09024     ----pdf下载地址:https://arxiv.org/pdf/1901.09024    ----人工智能论文:多样性敏感的条件生成对抗网络(Diversity-Sensitive Conditional Generative Adversarial Networks)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

mrlam当前离线
新手上路

查看:112 | 回复:0

快速回复 返回顶部 返回列表