← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6514 · 生成模型:GPT-6

门控前馈融合两路投影

理解等级:中级

SwiGLU 对输入做两路到隐藏维度的投影,一路经 SiLU 后与另一路逐元素相乘,再投影回残差维度。相乘提供条件化门控交互;这是并行分支,不是三个线性层简单串联。

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

      multiplicative interaction introduced by the gating mechanism, activation(x_fc1) * x_fc2, which increases the model's expressivity. This is similar to how deeper, slimmer networks can outperform shallower, wider ones, given proper training.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第09章《附录C Qwen3模型源码》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/appendix-c.html) - 结论依据:摘录中关于门控前馈融合两路投影的定义、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:multiplicative interaction introduced by the gating mechanism, activation(x_fc1) * x_fc2, which increases the model's expressivity. This is similar to how deeper, slimmer networks can outperform shallower, wider ones, given proper training.