Attention without RNN

Attention Layer: Attention without RNN

Compared with Simple RNN + Attention, it's not using hidden states from RNN to compute Key, Value, Query parameters.

• We study Seq2Seq model (encoder + decoder).

• Encoder’s inputs are vectors $x_1,x_2,...,x_m$.

• Decoder’s inputs are vectors $x'_1,x'_2,...,x'_t$.

• Keys and values are based on encoder’s inputs $x_1,x_2,...,x_m$.

  • Key: $k_{:i}=W_K x_i$

  • Value: $v_{:i}=W_V x_i$

• Queries are based on decoder’s inputs $x'_1,x'_2,...,x'_t$

• Query: $q_{:j}=W_Q x'_j$

• Compute weights: $\alpha_{:j}=Softmax(K^T q_{:j}) \in R^m$

• Context vector: $c_j=\alpha_{ij} v_{:1} + ... +\alpha_{mj} v_{:m}=V \alpha_{:j}$

  • Thus, $c_{:j}$is a function of $x'_j$and $[x_1,...x_m]$

• Output of attention layer: $C=[c_{:1}, c_{:2},c_{:3},...,c_{:t}]$

Attention layer summary

• Attention layer: $C=Attn(X,X')$

  • Encoder's inputs: $X=[x_1, x_2,...,x_m]$

  • Decoder's inputs: $X'=[x'_1,x'_2,...,x'_t]$

  • Parameters: $W_Q,W_K, W_V$

Self-Attention layer summary

• Self-Attention layer: $C=Attn(X,X)$

  • Inputs: $X=[x_1, x_2,...,x_m]$

  • Parameters: $W_Q,W_K, W_V$

Summary

• Attention was originally developed for Seq2Seq RNN models [1].

• Self-attention: attention for all the RNN models (not necessarily Seq2Seq models [2].

• Attention can be used without RNN [3].

• We learned how to build attention layer and self-attention layer.

Reference

• Original paper: Vaswani et al. Attention Is All You Need. In NIPS, 2017.

• youtube