残差连接Residual Connection
Residual Connection
解决什么问题What problem it solves
给网络留一条「直通通道」，让信号穿过很多层也不丢失。Gives the network a "pass-through lane" so the signal survives passing through many layers.
这是什么What it is
每一层除了输出加工后的结果，还把自己收到的原始输入原样加上去。
它来自图像识别里的 ResNet。
后来它成了 Transformer 能堆到几十层的前提之一。
Each layer adds its raw input back onto its processed output.
It came from ResNet in image recognition.
Later it became one of the prerequisites for Transformers stacking dozens of layers.
为什么需要知道Why it matters
网络越深越容易训不动——信号穿过几十层会走样。直通通道让模型「改一点、保留大部分」，深才可行。Deeper networks are harder to train — signals get distorted across dozens of layers. The pass-through lane lets the model "change a bit, keep most of it," which is what makes depth feasible.