层归一化Layer Normalization
Layer Normalization
解决什么问题What problem it solves
给网络每一层的输出做一次「调平」，让深层训练不失控。Levels out each layer's output so deep training doesn't spiral out of control.
这是什么What it is
对单个样本在某一层里的全部数值做归一化（拉回统一的均值和方差），再往下传。
Transformer 的每个子层后面都跟着它，和残差连接搭配使用。
Normalizes all values of a single sample within one layer (back to a common mean and variance) before passing them on.
In Transformers it follows every sub-layer, paired with residual connections.
为什么需要知道Why it matters
深层网络里数值容易越传越偏，一旦失控训练就崩。调平是「模型能稳定训到几十层」的幕后功臣之一。Values in deep networks drift as they propagate; once they spiral, training collapses. Leveling is one of the behind-the-scenes reasons models train stably at dozens of layers.
现在能做什么What you can do now
回到名词地图看第4站的其他 35 个词 →Back to the map: the other 35 terms in station 4 →