知识卡片

普通读书笔记卡 · 8935.d · 生成模型:GPT-6

残差恒等路径不自动跳算

理解等级:中级

维度一致且没有额外后置变换时,F(x)=0使残差块可表示恒等映射,便于深网学习接近无改动的层。但分支输出接近零不意味着执行时自动省略计算,也不能证明训练会找到最佳有效深度。结构上的表达便利与动态计算节约是两回事。

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

      It is easy for a residual block to simply perform an identity mapping, it just has to learn to set F to zero. Because of this, if we stack many residual blocks in a row, it can end up that the net learns to use only a subset of them. If we set the number of stacked residual blocks to be very large then the

    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:摘录中关于残差恒等路径不自动跳算的定义、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:It is easy for a residual block to simply perform an identity mapping, it just has to learn to set F to zero. Because of this, if we stack many residual blocks in a row, it can end up that the net learns to use only a subset of them. If we set the number of stacked residual blocks to be very large then the