知识卡片

普通读书笔记卡 · 8947.b · 生成模型:GPT-6

三角权重不自动阻止泄漏

理解等级:高级

即使值汇集只允许j,若查询、残差或其他路径仍包含正在预测的目标,模型仍可能读到答案。应明确输入与目标错位:位置i用可见前缀预测i+1,并让查询、键值、残差及归一化都遵守可见范围。三角形A只是整体因果设计的一部分。

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

      This kind of matrix is called causal because each output index i only depends on input indices j such that j < i. If A is an attention matrix, then this strategy is called causal attention. This is a masking strategy where each token can only attend to previous tokens in the sequence. This approach can dramatically speed up training because all the sub-sequence prediction

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第26章《Transformer》(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/c026.xhtml) - 结论依据:摘录中关于三角权重不自动阻止泄漏的定义、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:This kind of matrix is called causal because each output index i only depends on input indices j such that j < i. If A is an attention matrix, then this strategy is called causal attention. This is a masking strategy where each token can only attend to previous tokens in the sequence. This approach can dramatically speed up training because all the sub-sequence prediction