序列模型
定义:
![](https://img.haomeiwen.com/i14491816/bf863c1f1a74ac06.png)
类型:语音识别、情感分类、机器翻译
![](https://img.haomeiwen.com/i14491816/357711b18a536189.png)
![](https://img.haomeiwen.com/i14491816/6fb60b96309a70d5.png)
![](https://img.haomeiwen.com/i14491816/34999c12e09dd1f5.png)
为什么序列模型使用CNN效果不好:
![](https://img.haomeiwen.com/i14491816/a018300bd96eb051.png)
循环神经网络
定义:
![](https://img.haomeiwen.com/i14491816/92bfc0bcddf5a201.png)
类型:
![](https://img.haomeiwen.com/i14491816/d1bd59f9960a6349.png)
基础循环网络
![](https://img.haomeiwen.com/i14491816/5aa01e2f255974c9.png)
![](https://img.haomeiwen.com/i14491816/223fef452137ebb2.png)
![](https://img.haomeiwen.com/i14491816/a0ee1e9625bfb8fc.png)
词的表示
为了让网络理解我们的输入,需要用词进行向量表示
![](https://img.haomeiwen.com/i14491816/b090ee738e13e0ec.png)
![](https://img.haomeiwen.com/i14491816/a26ab1223b60560d.png)
![](https://img.haomeiwen.com/i14491816/89682667bb76922e.png)
![](https://img.haomeiwen.com/i14491816/f7fff4e5ee1b1194.png)
矩阵运算表示
![](https://img.haomeiwen.com/i14491816/6a4586564385dc20.png)
交叉熵损失
![](https://img.haomeiwen.com/i14491816/156522a81a0b2076.png)
时序反向传播算法(BPTT)
![](https://img.haomeiwen.com/i14491816/c2fa5aa813cc952a.png)
![](https://img.haomeiwen.com/i14491816/917ea314e4961379.png)
![](https://img.haomeiwen.com/i14491816/60d79f0496b040db.png)
梯度消失、梯度爆炸
![](https://img.haomeiwen.com/i14491816/f8cae96ef13cf246.png)
RNN总结
![](https://img.haomeiwen.com/i14491816/3244df4fe374065e.png)
![](https://img.haomeiwen.com/i14491816/c17d357c4a54ffda.png)
案例
流程:
![](https://img.haomeiwen.com/i14491816/756338ecd5557d25.png)
![](https://img.haomeiwen.com/i14491816/22d77f219009ee50.png)
单个cell的前向传播:
![](https://img.haomeiwen.com/i14491816/7286a1a12fc86aeb.png)
![](https://img.haomeiwen.com/i14491816/2263d00962094cb2.png)
所有cell的前向传播
![](https://img.haomeiwen.com/i14491816/524a729e9a3e403a.png)
![](https://img.haomeiwen.com/i14491816/c68ce6cfb47f3d40.png)
![](https://img.haomeiwen.com/i14491816/759a4447c88c16b0.png)
![](https://img.haomeiwen.com/i14491816/c7f96b2857780d13.png)
单个cell的反向传播
![](https://img.haomeiwen.com/i14491816/ce1063a834c2e1c1.png)
![](https://img.haomeiwen.com/i14491816/ad8c9945d4e6004a.png)
所有cell的反向传播(p86)
RNN的结构改进
GRU(门控循环单元)
![](https://img.haomeiwen.com/i14491816/c2ab55d5cc509d4e.png)
本质问题解决:
![](https://img.haomeiwen.com/i14491816/7795fa77d81a4a02.png)
LSTM(长短记忆网络)
![](https://img.haomeiwen.com/i14491816/15803ca28d4a9bc8.png)
作用是便于记忆更长距离的时间状态
网友评论