← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8998.a · 生成模型:GPT-6

条件前缀保留输出自回归

理解等级:中级

条件自回归分解为∏ip(yi|y,x),可以把x作为序列前缀,也可用交叉注意力等方式提供条件。训练目标只需明确哪些位置计算损失、各位置可读什么;直接建模完整[x,y]的联合似然,与只优化条件似然不是同一个目标。

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

      Autoregressive models (section 32.7) are already modeling a sequence of conditional distributions. So, to condition them on some inputs x, we can simply concatenate x as a prefix to the sequence of y values we are modeling, yielding the new sequence: [x1, …, xm, y1, …, yn]. The probability model factorizes as:

    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:摘录中关于条件前缀保留输出自回归的定义、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:Autoregressive models (section 32.7) are already modeling a sequence of conditional distributions. So, to condition them on some inputs x, we can simply concatenate x as a prefix to the sequence of y values we are modeling, yielding the new sequence: [x1, …, xm, y1, …, yn]. The probability model factorizes as: