2D-LSTM

LSTM

Understanding LSTM Networks人人都能看懂的LSTM 这两篇文章介绍了 LSTM 的原理。

2D-LSTM

2D-LSTM 是作用于三维输入( W × H × D W \times H \times D W×H×D )的 LSTM ,分别取横向和纵向上一时刻的隐藏状态和输出作为该时刻的输入,如下图所示
2D-LSTM
数据传播的顺序依靠对角线原则,如下图所示
2D-LSTM
图中的数字表示计算的顺序。
下图展示了 2D-LSTM 单元的结构,蓝线表示与标准单元不同的地方。
2D-LSTM
上图中 x j , i x_{j, i} xj,i​ 为当前的输入, s j , i − 1 s_{j, i-1} sj,i−1​ 为上一时刻横向的输出, s j − 1 , i s_{j-1, i} sj−1,i​ 为上一时刻纵向的输出。
input gate
i j , i = σ ( W 1 x j , i + U 1 s j − 1 , i + V 1 s j , i − 1 ) i_{j, i} = \sigma(W_1x_{j, i} + U_1s_{j-1, i} + V_1s_{j, i-1}) ij,i​=σ(W1​xj,i​+U1​sj−1,i​+V1​sj,i−1​)
output gate
o j , i = σ ( W 2 x j , i + U 2 s j − 1 , i + V 2 s j , i − 1 ) o_{j, i} = \sigma(W_2x_{j, i} + U_2s_{j-1, i} + V_2s_{j, i-1}) oj,i​=σ(W2​xj,i​+U2​sj−1,i​+V2​sj,i−1​)
candidate value
c ^ j , i = g ( W 3 x j , i + U 3 s j − 1 , i + V 3 s j , i − 1 ) \hat{c}_{j, i} = g(W_3x_{j, i} + U_3s_{j-1, i} + V_3s_{j, i-1}) c^j,i​=g(W3​xj,i​+U3​sj−1,i​+V3​sj,i−1​)
forget gate
f j , i = σ ( W 4 x j , i + U 4 s j − 1 , i + V 4 s j , i − 1 ) f_{j, i} = \sigma(W_4x_{j, i} + U_4s_{j-1, i} + V_4s_{j, i-1}) fj,i​=σ(W4​xj,i​+U4​sj−1,i​+V4​sj,i−1​)
2D-LSTM 新加入了一个系数,用于比较 s j − 1 , i s_{j-1, i} sj−1,i​ 和 s j , i − 1 s_{j, i-1} sj,i−1​ 的重要程度。
λ j , i = σ ( W 5 x j , i + U 5 s j − 1 , i + V 5 s j , i − 1 ) \lambda_{j, i} = \sigma(W_5x_{j, i} + U_5s_{j-1, i} + V_5s_{j, i-1}) λj,i​=σ(W5​xj,i​+U5​sj−1,i​+V5​sj,i−1​)
新状态
c j , i = f j , i ∘ [ λ j , i ∘ c j − 1 , i + ( 1 − λ j , i ) ∘ c j , i − 1 ] + c ^ j , i ∘ i j , i c_{j, i} = f_{j, i} \circ [\lambda_{j, i} \circ c_{j-1, i} + (1 - \lambda_{j, i}) \circ c_{j, i-1}] + \hat{c}_{j, i} \circ i_{j, i} cj,i​=fj,i​∘[λj,i​∘cj−1,i​+(1−λj,i​)∘cj,i−1​]+c^j,i​∘ij,i​
输出
s j , i = g ( c j , i ∘ o j , i ) s_{j, i} = g(c_{j, i} \circ o_{j, i}) sj,i​=g(cj,i​∘oj,i​)

Reference

[1] Bahar, P. , C. Brix , and H. Ney . “Towards Two-Dimensional Sequence to Sequence Model in Neural Machine Translation.” (2018).
[2] Voigtlaender, P. , P. Doetsch , and H. Ney . “Handwriting Recognition with Large Multidimensional Long Short-Term Memory Recurrent Neural Networks.” 2016 15th International Conference on Frontiers in Handwriting Recognition (ICFHR) IEEE, 2017.

上一篇:CSS复习高级部分(三)web字体、2D转换transfroms、3D转换


下一篇:CF3A Shortest path of the king