← 返回起始笔记位置

知识卡片

普通读书笔记卡 · 8986 · 生成模型:GPT-6

自回归似然是条件项之和

理解等级:中级

联合对数似然等于各位置真实值的条件对数概率之和,因此训练可以使用下一像素交叉熵。模型输入仅含目标之前的信息,不能包含当前真值或未来路径。真实前缀可使训练位置并行计算,生成时前缀则来自模型先前抽样。

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

      What about the third pixel, conditioned on the first two? Well, this is again a problem of the same form: a 256-way softmax regression conditioned on some observations. Now you can see the induction: modeling each next pixel in the sequence is a softmax regression problem that models p(xn | x1, …, xn–1). We show how the cross-entropy loss can be computed in figure 32.11.

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第32章《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/c032.xhtml) - 结论依据:摘录中关于自回归似然是条件项之和的定义、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:What about the third pixel, conditioned on the first two? Well, this is again a problem of the same form: a 256-way softmax regression conditioned on some observations. Now you can see the induction: modeling each next pixel in the sequence is a softmax regression problem that models p(xn | x1, …, xn–1). We show how the cross-entropy loss can be computed in figure 32.11.