update lstm&gru

Section06-Neural-Machine-Translation/section06.tex

@@ -1565,15 +1565,15 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
         \end{scope}
 
         \begin{scope}
-            \node[wordnode,anchor=south] () at ([xshift=0.5\base]aux21) {$h_t$};
+            \node[wordnode,anchor=south] () at ([xshift=0.5\base]aux21) {$h_{t-1}$};
             \node[wordnode,anchor=west] () at (aux12) {$x_t$};
-            \node[wordnode,anchor=south] () at ([xshift=0.5\base]aux51) {$c_t$};
+            \node[wordnode,anchor=south] () at ([xshift=0.5\base]aux51) {$c_{t-1}$};
             \visible<3->{
-                \node[wordnode,anchor=south] () at ([xshift=-0.5\base]aux59) {$c_{t+1}$};
+                \node[wordnode,anchor=south] () at ([xshift=-0.5\base]aux59) {$c_{t}$};
             }
             \visible<4->{
-                \node[wordnode,anchor=east] () at (aux68) {$h_{t+1}$};
-                \node[wordnode,anchor=south] () at ([xshift=-0.5\base]aux29) {$h_{t+1}$};
+                \node[wordnode,anchor=east] () at (aux68) {$h_{t}$};
+                \node[wordnode,anchor=south] () at ([xshift=-0.5\base]aux29) {$h_{t}$};
             }
         \end{scope}
 
@@ -1584,27 +1584,27 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
         \begin{scope}
             \visible<1->{
                 % forget gate formula
-                \node[formulanode,anchor=south east,text width=3.4cm] () at ([shift={(4\base,1.5\base)}]aux51) {遗忘门\\$f_t=\sigma(W_f[h_t,x_t]+b_f)$};
+                \node[formulanode,anchor=south east,text width=3.4cm] () at ([shift={(4\base,1.5\base)}]aux51) {遗忘门\\$f_t=\sigma(W_f[h_{t-1},x_t]+b_f)$};
             }
             \visible<2->{
                 % input gate formula
-                \node[formulanode,anchor=north east] () at ([shift={(4\base,-1.5\base)}]aux21) {输入门\\$i_t=\sigma(W_i[h_t,x_t]+b_i)$\\$\hat{c}_t=\mathrm{tanh}(W_c[h_t,x_t]+b_c)$};
+                \node[formulanode,anchor=north east] () at ([shift={(4\base,-1.5\base)}]aux21) {输入门\\$i_t=\sigma(W_i[h_{t-1},x_t]+b_i)$\\$\hat{c}_t=\mathrm{tanh}(W_c[h_{t-1},x_t]+b_c)$};
             }
             \visible<3->{
                 % cell update formula
-                \node[formulanode,anchor=south west,text width=3.02cm] () at ([shift={(-4\base,1.5\base)}]aux59) {记忆更新\\$c_{t+1}=f_t\cdot c_t+i_t\cdot \hat{c}_t$};
+                \node[formulanode,anchor=south west,text width=3.02cm] () at ([shift={(-4\base,1.5\base)}]aux59) {记忆更新\\$c_{t}=f_t\cdot c_{t-1}+i_t\cdot \hat{c}_t$};
             }
             \visible<4->{
                 % output gate formula
-                \node[formulanode,anchor=north west] () at ([shift={(-4\base,-1.5\base)}]aux29) {输出门\\$o_t=\sigma(W_o[h_t,x_t]+b_o)$\\$h_{t+1}=o_t\cdot \mathrm{tanh}(c_{t+1})$};
+                \node[formulanode,anchor=north west] () at ([shift={(-4\base,-1.5\base)}]aux29) {输出门\\$o_t=\sigma(W_o[h_{t-1},x_t]+b_o)$\\$h_{t}=o_t\cdot \mathrm{tanh}(c_{t})$};
             }
         \end{scope}
     \end{tikzpicture}
 \end{center}
 
 {\scriptsize\begin{tabular}{l}
-    *$x_t$: 前一层的输出，$h_t$: 同一层上一时刻的隐藏状态\\
-    *$c_t$: 同一层上一时刻的记忆
+    *$x_t$: 上一层的输出，$h_{t-1}$: 同一层上一时刻的隐藏状态\\
+    *$c_{t-1}$: 同一层上一时刻的记忆
 \end{tabular}}
 \end{frame}
 
@@ -1735,11 +1735,11 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
         \end{scope}
 
         \begin{scope}
-            \node[wordnode,anchor=south] () at (aux71) {$h_t$};
+            \node[wordnode,anchor=south] () at (aux71) {$h_{t-1}$};
             \node[wordnode,anchor=west] () at (aux12) {$x_t$};
             \visible<3->{
-                \node[wordnode,anchor=east] () at (aux87) {$h_{t+1}$};
-                \node[wordnode,anchor=south] () at (aux78) {$h_{t+1}$};
+                \node[wordnode,anchor=east] () at (aux87) {$h_{t}$};
+                \node[wordnode,anchor=south] () at (aux78) {$h_{t}$};
             }
         \end{scope}
 
@@ -1750,15 +1750,15 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
         \begin{scope}
             \visible<1->{
                 % reset gate formula
-                \node[formulanode,anchor=west,text width=4cm] (reset) at ([shift={(\base,0.7\base)}]aux78) {重置门\\$r_t=\sigma(W_r[h_t,x_t])$};
+                \node[formulanode,anchor=west,text width=4cm] (reset) at ([shift={(\base,0.7\base)}]aux78) {重置门\\$r_t=\sigma(W_r[h_{t-1},x_t])$};
             }
             \visible<2->{
                 % update gate formula
-                \node[formulanode,anchor=north west,text width=4cm] (update) at ([yshift=-0.5\base]reset.south west) {更新门\\$u_t=\sigma(W_u[h_t,x_t])$};
+                \node[formulanode,anchor=north west,text width=4cm] (update) at ([yshift=-0.5\base]reset.south west) {更新门\\$u_t=\sigma(W_u[h_{t-1},x_t])$};
             }
             \visible<3->{
                 % hidden update formula
-                \node[formulanode,anchor=north west,text width=4cm] () at ([yshift=-0.5\base]update.south west) {隐藏状态更新\\$\hat{h}_t=\mathrm{tanh}(W[r_t\cdot h_t,x_t])$\\$h_{t+1}=(1-u_t)\cdot h_t+u_t\cdot \hat{h}_t$};
+                \node[formulanode,anchor=north west,text width=4cm] () at ([yshift=-0.5\base]update.south west) {隐藏状态更新\\$\hat{h}_t=\mathrm{tanh}(W[r_t\cdot h_{t-1},x_t])$\\$h_{t}=(1-u_t)\cdot h_{t-1}+u_t\cdot \hat{h}_t$};
             }
         \end{scope}
     \end{tikzpicture}
@@ -1766,7 +1766,7 @@ NLP问题的隐含结构假设 & 无隐含结构假设，端到端学习 \\
 
 {\footnotesize\begin{tabular}{l}
     *$x_t$: 上一层的输出\\
-    *$h_t$: 同一层上一时刻的隐藏状态
+    *$h_{t-1}$: 同一层上一时刻的隐藏状态
 \end{tabular}}
 \end{frame}