← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8989 · 生成模型:GPT-6

编码与生成可以共同学习

理解等级:中级

编码把数据映射到表示,生成从表示构造数据,两个方向可通过重建和概率约束相互帮助。但压缩、多对一及随机观测使它们通常不能成为全空间互逆函数。“相反方向”是建模思路,不是所有生成网络都能直接逐层求逆。

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

      The intuition is that generative models map a simple base distribution (i.e., noise) to data, whereas representation learning maps data to simple underlying representations (embeddings). These two problems are, essentially, inverses of each other. Many algorithms explicitly treat them as inverse problems, where solving the problem in one direction can inform the solution in the other direction.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第33章《生成建模与表示学习》(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/c033.xhtml) - 结论依据:摘录中关于编码与生成可以共同学习的定义、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 intuition is that generative models map a simple base distribution (i.e., noise) to data, whereas representation learning maps data to simple underlying representations (embeddings). These two problems are, essentially, inverses of each other. Many algorithms explicitly treat them as inverse problems, where solving the problem in one direction can inform the solution in the other direction.