序列模型
吴恩达 Andrew Ng

Sequence models & Attention mechanism

Various sequence to sequence architectures

Basic Models

• sequence to sequence model
• image captioning 图像描述

Picking the most likely sentence

• condition language model

  

• just pick one word at a time (greedy search) is not always optimal

• approximate search algorithm

Beam Search 集束搜索

• coalition 编码，decodlition 解码

• beam width (B)集束宽，候选词的个数

• record top B possiblities of sentences

• 步骤
  
  
  

Refinements to Beam Search

• 1Tαy∑Tyy=1logP(y<t>|x,y<1>,⋯,y<t−1>)$\frac{1}{T_y^{\alpha }}\sum _{y=1}^{T_y}\log P(y^{<t>}|x,y^{<1>},\cdots ,y^{<t-1>})$

  • length normalization 长度归一化
    

  • numerical underflow 数值下溢，rounding error 四舍五入的误差

  • α$\alpha$ 是超参数

  • normalized log likelihood objective 归一化的对数似然目标函数

• large B: better result but computationally slower

• small B: worse result but faster

• Beam Search is not guaranteed to find exact maximum for argmaxyP(y|x)$arg\underset{y}{max}P(y|x)$

Bleu Score (optional)

• give a score to measure how good is the machine translation
• Bleu score on unigrams (一元词组), bigrams (二元词组)
• BP: brevity penalty 简短惩罚
• single real number evaluation metric 单一实数评估指标

Attention Model Intuition （注意力模型）

• 神经网络很难记忆长句子

• 一部分一部分来机器翻译

• 

Attention Model

Speech recognition - Audio data

Speech recognition

• ​false blank outputs 伪空白输出

• phonemes, hand-engineered basic units of cells

• end-to-end network, input an audio clip and directly output a transcript

• Connectionist Temporal Classification cost function

  collapse repeated characters not separated by “blank”

Trigger Word Detection 触发字检测

label​

Conclusion

Conclusion and thank you

Make the world a better place.​