← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8935.c · 生成模型:GPT-6

残差网络加法与拼接不同

理解等级:中级

ResNet块计算y=x+F(x),让分支学习相对输入的修正;U-Net常沿通道拼接,两者改变的张量结构不同。相加要求维度一致,若使用Wx投影作捷径,则映射为y=F(x)+Wx,F=0时得到Wx,并不必然是恒等映射。

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

      In the context of ResNets, skip connections are called residual connections. This kind of skip connection is added to the output of a block of layers F: (24.28) In a residual block like this, you can think of F as a residual that additively perturbs xin to transform it into an improved xout. If xout does not have the same dimensionality as xin, then we

    参考来源

    - 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第24章《卷积神经网络》(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/c024b.xhtml) - 结论依据:摘录中关于残差网络加法与拼接不同的定义、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:In the context of ResNets, skip connections are called residual connections. This kind of skip connection is added to the output of a block of layers F: (24.28) In a residual block like this, you can think of F as a residual that additively perturbs xin to transform it into an improved xout. If xout does not have the same dimensionality as xin, then we