Long Short-Term Memory (LSTM)

Motivation

• Memory cell
• Inputs are “commited” into memory. Later inputs “erase” early inputs
• An additional memory “cell” for long term memory
• Also being read and write from the current step, but less affected like 𝐻

LSTM Operations

• Forget gate
• Input Gate
• Candidate Content
• Output Gate

Forget

• Forget: remove information from cell $C$

• What to forget depends on:

  • the current input $X_t$
  • the previous memory $H_{t-1}$

• Forget gate: controls what should be forgotten $$ F_{t}=\operatorname{sigmoid}\left(W^{F X} X_{t}+W^{F H} H_{t-1}+b^{F}\right) $$

• Content to forget: $$ C = F_t * C_{t-1} $$

  • $F_{t,j}$ near 0: Forgetting the content stored in $C_j$

Write

• Write: Adding new information for cell $C$

• What to forget depends on:

  • the current input $X_t$
  • the previous memory $H_{t-1}$

• Input gate: controls what should be add $$ I_{t}=\operatorname{sigmoid}\left(W^{I X} X_{t}+W^{I H} H_{t-1}+b^{I}\right) $$

• Content to write: $$ \tilde{C}_{t}=\tanh \left(W^{C X} X_{t}+W^{C H} H_{t-1}+b^{I}\right) $$

• Write content: $$ C=C_{t-1}+I_{t} * \tilde{C}_{t} $$

Output

• Output: Reading information from cell 𝐶 (to store in the current state $H$)

• How much to write depends on:

  • the current input $X_t$
  • the previous memory $H_{t-1}$

• Forget gate: controls what should be output $$ O_{t}=\operatorname{sigmoid}\left(W^{OX} X_{t}+W^{OH} H_{t-1}+b^{O}\right) $$

• New state $$ H_t = O_t * \operatorname{tanh}(C_t) $$

LSTM Gradients

Truncated Backpropagation

What happens if the sequence is really long (E.g. Character sequences, DNA sequences, video frame sequences …)? $\to$ Back-propagation through time becomes exorbitant at large $T$

Solution: Truncated Backpropagation

• Divide the sequences into segments and truncate between segments
• However the memory is kept to remain some information about the past (rather than resetting)

TDNN vs. LSTM

	TDNN	LSTM
	For $t = 1,\dots, T$: $H_{t}=\mathrm{f}\left(\mathrm{W}_{1} \mathrm{I}_{\mathrm{t}-\mathrm{D}}+\mathrm{W}_{2} \mathrm{I}_{\mathrm{t}-\mathrm{D}+1}+\mathrm{W}_{3} \mathrm{I}_{\mathrm{t}-\mathrm{D}+2}+\cdots+W_{D} I_{t}\right)$	For $t = 1,\dots, T$: $H_{t}=f\left(W^{I} I_{t}+W^{H} H_{t-1}+b\right)$
Weights are shared over time?	Yes	Yes
Handle variable length sequences	Increasing long-range dependency learning by increasing $D$ Increasing $D$ also increases number of parameters	Can flexibly adapt to variable length sequences without changing structures
Gradient vanishing or exploding problem?	No	Yes
Parallelism?	Can be parallelized into 𝑂(1) (Assuming Matrix multiplication cost is O(1) thanks to GPUs…)	Sequential computation (because $𝐻_𝑇$ cannot be computed before $𝐻_{𝑇−1}$)