知识卡片

普通读书笔记卡 · 9012.b · 生成模型:GPT-6

无配对翻译不能保证类对齐

理解等级:高级

算法 37.2 先训练源预测器 f,再用对抗目标训练目标→源翻译器 g,推理为 F=f∘g。判别器只识别来源外观,不能阻止类别被错误映射;循环一致性可添加结构约束,但也不能唯一识别正确语义映射。训练 GAN 不是得到可靠域适配的充分条件。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Algorithm 37.2: Unpaired domain adaptation via adversarial translation. Here we use a GAN to translate from one domain to the other. This algorithm could be improved by adding cycle-consistency constraints (see CycleGAN from chapter 34, [531]) as was done in [211].

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第37章《迁移学习与适配》(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/c037.xhtml) - 结论依据:摘录中关于无配对翻译不能保证类对齐的定义、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:Algorithm 37.2: Unpaired domain adaptation via adversarial translation. Here we use a GAN to translate from one domain to the other. This algorithm could be improved by adding cycle-consistency constraints (see CycleGAN from chapter 34, [531]) as was done in [211].