koryako的博客

向前传播的是值，反向传播的是梯度 // every Unit corresponds to a wire in the diagrams var Unit = function(value, grad) { // value computed in the forward pass this.value = value; // the derivative

导数的链式法则的理解 var x = -2, y = 5,z=-4,h = 0.1;// whenever value of notation h was changeed, but the following derivateive value was not changed. var i=1; while (i<=10) {

1. 首先是概念， // 格式化节点，每个节点中包含的计算值和梯度值 function Unit(value, grad) { // 这个值是正向传播的值 this.value = value; // 这个值是反向传播的值， this.grad = grad; } function multiplyGate(){ };//乘法门 multipl

以下是对这边paper的翻译，仅当练手  DenseCap: Fully Convolutional Localization Networks for Dense Captioning We introduce the dense captioning task, which requires a computer vision system to b