知识卡片
只有仿射层时加深不增非线性
若连续层都为仿射变换,组合后仍可折叠为一个仿射变换。增加这种层数会改变参数化,却不会凭空形成非线性决策边界。激活函数等非线性操作因此承担重要表达职责,不能把矩阵乘法次数直接视为函数复杂度。
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
- · 神经网络
原书出处 · 1 处
- · 查看原书依据
the composition of N linear functions is a linear function. Activation layers add nonlinearity. Activation layers are typically pointwise functions, applying a scalar to scalar mapping on each dimension of the input vector. Typically parameters of these layers, if any, are not learned (but they can be). Some common activation layers are defined below and are plotted in figure 12.16:
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第12章《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/c012.xhtml)
- 结论依据:摘录中关于只有仿射层时加深不增非线性的定义、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:the composition of N linear functions is a linear function. Activation layers add nonlinearity. Activation layers are typically pointwise functions, applying a scalar to scalar mapping on each dimension of the input vector. Typically parameters of these layers, if any, are not learned (but they can be). Some common activation layers are defined below and are plotted in figure 12.16: