Skip to content

Commit

Permalink
Fix bad math alignment
Browse files Browse the repository at this point in the history
  • Loading branch information
Atcold committed Sep 15, 2020
1 parent df0ceb8 commit 30845e1
Showing 1 changed file with 7 additions and 4 deletions.
11 changes: 7 additions & 4 deletions docs/zh/week06/06-3.md
Original file line number Diff line number Diff line change
Expand Up @@ -171,9 +171,9 @@ Yann在图示的网络神经之间添加了一些特殊图形,用来表示从



## [基于时间的反向传播算法(Back Propagation through time)](https://www.youtube.com/watch?v=8cAffg2jaT0&t=855s)
## [基于时间的反向传播算法(Back Propagation through time)](https://www.youtube.com/watch?v=8cAffg2jaT0&t=855s)

### <!--Model architecture-->
### <!--Model architecture-->

### 模型架构

Expand All @@ -189,6 +189,7 @@ Yann在图示的网络神经之间添加了一些特殊图形,用来表示从
<!--Hidden representations are stated as-->

隐藏层表达式为:

$$
\begin{aligned}
\begin{cases}
Expand All @@ -206,6 +207,7 @@ $$
<!--The first equation indicates a non-linear function applied on a rotation of a stack version of input where the previous configuration of the hidden layer is appended. At the beginning, $h[0]$ is set 0. To simplify the equation, $W_h$ can be written as two separate matrices, $\left[ W_{hx}\ W_{hh}\right]$, thus sometimes the transformation can be stated as-->

$h[t]$是一个非线性函数,将堆叠形式的输入进行旋转操作,其中加入了前一步的隐藏层。$h[0]$初始值设为1。$W_h$可简化为两个不同的矩阵$\left[ W_{hx}\ W_{hh}\right]$,变换后的公式可以写成

$$
W_{hx}\cdot x[t]+W_{hh}\cdot h[t-1]
$$
Expand Down Expand Up @@ -233,6 +235,7 @@ $y[t]$由最后一步的旋转计算得出,之后我们可以用链式法则
<!--If BPTT period $T$ is set to 3, the first input $x[1:T]$ and output $y[1:T]$ for RNN is determined as-->

如果将BPTT的周期$T$设为3, 那么第一组RNN的输入$x[1:T]$和输出$y[1:T]$为:

$$
\begin{aligned}
x[1:T] &= \begin{bmatrix}
Expand Down Expand Up @@ -261,7 +264,7 @@ $$

## 梯度消失与梯度爆炸(Vanishing and Exploding Gradient)

### <!--Problem-->
### <!--Problem-->

### 问题

Expand All @@ -278,7 +281,7 @@ $$

在经典RNN中,梯度会通过所有可能的箭头来传播。这有很大的几率会使梯度爆炸或者消失。例如,梯度在时刻1的时候非常大(图中白色圆点)。当进行一个旋转后,梯度在时刻3收缩,即消失。

### <!--Solution-->
### <!--Solution-->

### 解决办法

Expand Down

0 comments on commit 30845e1

Please sign in to comment.