人工智能培训

搜索

人工智能论文:更深入地了解对抗性损失(Towards a Deeper Understanding of Adversarial Losses)

[复制链接]
usjerseys889 发表于 2019-1-28 11:20:34 | 显示全部楼层 |阅读模式
usjerseys889 2019-1-28 11:20:34 100 0 显示全部楼层
人工智能论文:更深入地了解对抗性损失(Towards a Deeper Understanding of Adversarial Losses)最近的工作提出了各种对抗性损失,用于训练生殖对抗性网络。然而,目前尚不清楚某些类型的功能是有效的对抗性损失功能,以及这些功能如何相互影响。在本文中,我们旨在通过解耦其组件功能和规范化术语的影响来更深入地了解对抗性损失。我们首先得出组件函数的一些必要和充分条件,使得对抗性损失是数据和模型分布之间的类似分歧的测量。为了系统地比较不同的对抗性损失,我们然后提出DANTest,一种基于判别性对抗性网络的新的简单框架。通过这种框架,我们通过结合不同的组件功能和正则化方法来评估一系列广泛的对抗性损失。这项研究引出了一些关于对抗性损失的新见解。为了重现性,可以通过此https URL获得所有源代码。
Recent work has proposed various adversarial losses for training generativeadversarial networks.Yet, it remains unclear what certain types of functionsare valid adversarial loss functions, and how these loss functions performagainst one another.In this paper, we aim to gain a deeper understanding ofadversarial losses by decoupling the effects of their component functions andregularization terms.We first derive some necessary and sufficient conditionsof the component functions such that the adversarial loss is a divergence-likemeasure between the data and the model distributions.In order tosystematically compare different adversarial losses, we then propose DANTest, anew, simple framework based on discriminative adversarial networks.With thisframework, we evaluate an extensive set of adversarial losses by combiningdifferent component functions and regularization approaches.This study leadsto some new insights into the adversarial losses.For reproducibility, allsource code is available at this https URL .人工智能论文:更深入地了解对抗性损失(Towards a Deeper Understanding of Adversarial Losses) pMuF44YyFKkY1ifu.jpg
URL地址:https://arxiv.org/abs/1901.08753     ----pdf下载地址:https://arxiv.org/pdf/1901.08753    ----人工智能论文:更深入地了解对抗性损失(Towards a Deeper Understanding of Adversarial Losses)
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

usjerseys889当前离线
新手上路

查看:100 | 回复:0

快速回复 返回顶部 返回列表