new updates

Section05-Neural-Networks-and-Language-Modeling/section05-test.tex

@@ -120,41 +120,11 @@
 
 
 %%%------------------------------------------------------------------------------------------------------------
-%%% 循环神经网络
-\begin{frame}{循环神经网络(Recurrent Neural Networks)}
+%%% 循环神经网络的结构
+\begin{frame}{循环单元}
 
 \begin{itemize}
-\item FNN LM固然有效，但是和传统的$n$-gram LM一样，需要依赖\alert{有限上下文}假设
-\begin{center}
-\begin{tikzpicture}
-\begin{scope}
-\node [anchor=west] (w0) at (0,0) {$w_1$};
-\node [anchor=west] (w1) at ([xshift=0.5em]w0.east) {$w_2$};
-\node [anchor=west] (w2) at ([xshift=0.5em]w1.east) {$...$};
-\node [anchor=west] (w3) at ([xshift=0.5em]w2.east) {$w_{m-n+1}$};
-\node [anchor=west] (w4) at ([xshift=0.5em]w3.east) {$...$};
-\node [anchor=west,fill=green!20!white] (w5) at ([xshift=0.5em]w4.east) {$w_{m}$};
-\draw [->,thick,ublue] (w5.south).. controls +(210:0.5) and +(-30:0.5) .. (w3.south);
-\draw [->,thick,red] (w5.north).. controls +(150:1) and +(30:1) .. (w1.north);
-\draw [->,very thick,ublue] ([xshift=-5em,yshift=1em]w0.west) -- ([xshift=-6.5em,yshift=1em]w0.west) node [pos=0,right] {\scriptsize{依赖}};
-\draw [->,very thick,red] ([xshift=-5em,yshift=-0.5em]w0.west) -- ([xshift=-6.5em,yshift=-0.5em]w0.west) node [pos=0,right] {\scriptsize{不依赖}};
-
-\end{scope}
-\end{tikzpicture}
-\end{center}
-\item<2-> 能否直接对原始问题建模，即定义函数$g$，对于任意的$w_{1} ... w_{m}$有
-    \vspace{-0.5em}
-    \begin{displaymath}
-    g(w_{1} ... w_{m}) \approx \textrm{P}(w_m | w_{1} ... w_{m-1})
-    \end{displaymath}
-\item<3-> \textbf{循环神经网络(RNNs)}可以很好的解决上述问题，因此也被成功的应用于语言建模任务
-	\begin{itemize}
-	\item 它假设每个词的生成都依赖已经生成的所有词
-	\item 对于不同位置的词的生成概率都可以用同一个函数描述
-	\end{itemize}
-	
-        \textbf{Recurrent Neural Network Based Language Model}\\
-        \textbf{Mikolov et al., 2010, In Proc. of Interspeech, 1045-1048}
+\item 假设有输入序列$(\textbf{x}_0,\textbf{x}_1,...,\textbf{x}_t,...)$，这里$\textbf{x}_t$表示序列中第$t$个元素，也被称作时刻$t$所对应的输入。它所对应的输出序列是$(\textbf{y}_0,\textbf{y}_1,...,\textbf{y}_t,...)$。 循环神经网络的核心是`` 记忆''任意长时间的历史
 \end{itemize}
 
 \end{frame}