循环神经网络层

目录

• Recap
• Sentiment Analysis
  • Proposal
• S1.Weight sharing
  • Naive version
  • Weight share
• S2.Consistent memory
• Unfolded model
• Formulation
• Overall Diagram
• One more thing
• How to Train?

Recap

Sentiment Analysis

Proposal

• Long sentence

  • 100 + words
  • too much parameters[wN,bN$w_N,b_N$]

• No context information

  • consistent tensor

S1.Weight sharing

• 类似于卷积的卷积核，卷积视野，权重分享

Naive version

Weight share

• 权重分享

S2.Consistent memory

• 需要一个东西存储语境，也就是每个单词的语境信息

• 不断地对语境信息进行堆叠得到h5，直接使用h5作为判断标准

Unfolded model

• 在时间轴上折叠，不断地更新h

Formulation

• 通过激活函数tanh不断地叠加上个时间戳的信息

Overall Diagram

• 全览分解图

One more thing

How to Train?