知识卡片
前馈与残差归一保留不同职责
注意力混合位置,逐位置前馈加入非线性变换,残差保留通路,LayerNorm按特征规范尺度。LayerNorm不把所有值压到0至1;原始Post-LN与其他Norm布局也需按公式区分,不能混称普通缩放。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
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
9.2.2 Connecting all the pieces Now that you’ve seen the hows and whys of BPE, embeddings, positional encoding, and multi-head self-attention, you understand all the elements of a transformer layer. You just need a lower-dimensional linear layer at the output to collect all those attention weights together to create the output sequence of embeddings, and the linear layer output needs to be scaled (normalized) so
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:摘录中关于前馈与残差归一保留不同职责的定义、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:9.2.2 Connecting all the pieces Now that you’ve seen the hows and whys of BPE, embeddings, positional encoding, and multi-head self-attention, you understand all the elements of a transformer layer. You just need a lower-dimensional linear layer at the output to collect all those attention weights together to create the output sequence of embeddings, and the linear layer output needs to be scaled (normalized) so