← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6745.a · 生成模型:GPT-6

解码器第三子层是交叉注意力

理解等级:中级

原始编码器解码器Transformer的解码块含因果自注意、对源编码的交叉注意和前馈。因果mask限制第一项读取未来目标,不是新增的零一全连接子层;交叉注意也通常出现在各解码块,而非只第一层。

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

      Decoders The decoder is nearly identical to the encoder but has three sublayers instead of two. The new sublayer is a fully connected layer similar to the multi-head self-attention matrix, but the new sublayer contains only zeros and ones. This creates a masking of the output sequences to the right of the current target token (in a left-to-right language, like English). This ensures that predictions

    参考来源

    - 位置:《Natural Language Processing in Action, Second Edition》第09章《可堆叠的Transformer》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Natural Language Processing in Action, Second Edition/_epub-src/OEBPS/Text/chapter-9.html) - 结论依据:摘录中关于解码器第三子层是交叉注意力的定义、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:Decoders The decoder is nearly identical to the encoder but has three sublayers instead of two. The new sublayer is a fully connected layer similar to the multi-head self-attention matrix, but the new sublayer contains only zeros and ones. This creates a masking of the output sequences to the right of the current target token (in a left-to-right language, like English). This ensures that predictions