知识卡片
匹配边缘分布不能确定映射
把X样本映射到Y域并匹配p(Y),不要求每个输入对应特定输出。许多置换或语义错误的映射都可保持同一输出分布。对抗目标因此只解决部分约束,不能单独识别所期望的图像翻译关系,也不必使映射一一对应。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 条件生成模型
原书出处 · 1 处
- · 查看原书依据
Such a GAN may indeed learn the correct mapping from x to y values—that's one of the solutions that minimizes the loss—but there are many symmetries in this learning problem, that is, many different mappings achieve equivalent loss. For example, consider that we permute the mapping: if {x(i), y(i)} is the true mapping then consider we instead recover {x(i), y(i+1)} after learning (circularly shifting at
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第34章《条件生成模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c034.xhtml)
- 结论依据:摘录中关于匹配边缘分布不能确定映射的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Such a GAN may indeed learn the correct mapping from x to y values—that's one of the solutions that minimizes the loss—but there are many symmetries in this learning problem, that is, many different mappings achieve equivalent loss. For example, consider that we permute the mapping: if {x(i), y(i)} is the true mapping then consider we instead recover {x(i), y(i+1)} after learning (circularly shifting at