← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8986.a · 生成模型:GPT-6

重叠前缀允许复用计算

理解等级:中级

一批相邻的下一项预测具有大量共同前缀,在满足因果约束的模型中可以共享中间表示,提高训练效率。但滑动窗口、位置变化和截断上下文会影响缓存有效性,不能任意把不同窗口的表示当作完全相同;仍需检查目标移位与掩码。

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

      that is, the example sequences overlap. This second way can be much more efficient. The reason is because in order to predict xn from x1, . . . , xn–1, we typically have to compute representations of x1, . . . , xn–1, and these same representations can be reused for predicting the next item over, that is, xn+1 from x2, . . . ,

    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:摘录中关于重叠前缀允许复用计算的定义、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:that is, the example sequences overlap. This second way can be much more efficient. The reason is because in order to predict xn from x1, . . . , xn–1, we typically have to compute representations of x1, . . . , xn–1, and these same representations can be reused for predicting the next item over, that is, xn+1 from x2, . . . ,