← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8947.a · 生成模型:GPT-6

因果掩码应作用于分数

理解等级:中级

自回归预测需要阻断对未来信息的读取。常见实现是在Softmax前将不可见位置分数设为负无穷,再对可见键归一化;直接将归一化后的权重清零会改变总权重。若某行全部不可见,需要定义初始词元或明确处理,不能依赖未定义的归一化结果。

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

      Sometimes we want to restrict which tokens can attend to which. This can be done by masking the attention matrix, which just means fixing some of the weights to be zero. This can be useful for many settings, including learning features via masked autoencoding [192], cross-attention between different data modalities [497], and for sequential prediction [78]. To illustrate, we will describe the sequential prediction use

    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:摘录中关于因果掩码应作用于分数的定义、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:Sometimes we want to restrict which tokens can attend to which. This can be done by masking the attention matrix, which just means fixing some of the weights to be zero. This can be useful for many settings, including learning features via masked autoencoding [192], cross-attention between different data modalities [497], and for sequential prediction [78]. To illustrate, we will describe the sequential prediction use