知识卡片
循环一致仍有配对歧义
若一个错误但可逆的X到Y映射保持目标分布,并以相反映射还原,两个循环损失仍可很小。因此往返恢复不证明内容语义正确,也不自动恢复真实配对。需要结构偏好、额外信息或任务评估消除剩余歧义。
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
The cycle-consistency losses encourage that the translation is a one-to-one function, and thereby reduce the number of valid solutions. This is a good idea when the true solution is indeed one-to-one, as is often the case in translation problems. However, other symmetries still exist and may confound methods like CycleGAN: if we permute the correct X-to-Y mapping and apply the inverse permutation to the Y-to-X
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:摘录中关于循环一致仍有配对歧义的定义、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:The cycle-consistency losses encourage that the translation is a one-to-one function, and thereby reduce the number of valid solutions. This is a good idea when the true solution is indeed one-to-one, as is often the case in translation problems. However, other symmetries still exist and may confound methods like CycleGAN: if we permute the correct X-to-Y mapping and apply the inverse permutation to the Y-to-X