知识卡片
层归一化作用于词元特征
此处LayerNorm分别沿每个词元的特征维归一化,通常加入数值稳定项,并可包含可学习缩放和平移参数。它不应混入未来位置的统计量。书中“没有可学习参数”的描述只适用于关闭仿射参数的特定实现。
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
This architecture uses layer normalization (section 12.7.3) before each attention layer and before each token-wise MLP layer. The normalization is done within each token (the token code vector is treated as a akin to a layer; each dimension of this vector is standardized by the mean and variance over all dimensions of this vector), so in figure 12.7.3 we refer to this layer as token
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:摘录中关于层归一化作用于词元特征的定义、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 architecture uses layer normalization (section 12.7.3) before each attention layer and before each token-wise MLP layer. The normalization is done within each token (the token code vector is treated as a akin to a layer; each dimension of this vector is standardized by the mean and variance over all dimensions of this vector), so in figure 12.7.3 we refer to this layer as token