stem/AI/Neural Networks/Transformers/Attention.md

- Meant to mimic cognitive attention
	- Picks out relevant bits of information
	- Use gradient descent
- Used in 90s
	- Multiplicative modules
	- Sigma pi units
	- Hyper-networks
- Draw from relevant state at any preceding point along sequence
	- Addresses [[RNN]]s vanishing gradient issues
	- [[LSTM]] tends to poorly preserve far back [[Neural Networks#Knowledge|knowledge]]
- Attention layer access all previous states and weighs according to learned measure of relevance
	- Allows referring arbitrarily far back to relevant tokens
- Can be addd to [[RNN]]s
- In 2016, a new type of highly parallelisable _decomposable attention_ was successfully combined with a [[Architectures|feedforward]] network
	- Attention useful in of itself, not just with [[RNN]]s
- [[Transformers]] use attention without recurrent connections
	- Process all tokens simultaneously
	- Calculate attention weights in successive layers

# Scaled Dot-Product
- Calculate attention weights between all tokens at once
- Learn 3 [[Weight Init|weight]] matrices
	- Query
		- $W_Q$
	- Key
		- $W_K$
	- Value
		- $W_V$
- Word vectors
	- For each token, $i$, input word embedding, $x_i$
		- Multiply with each of above to produce vector
	- Query Vector
		- $q_i=x_iW_Q$
	- Key Vector
		- $k_i=x_iW_K$
	- Value Vector
		- $v_i=x_iW_V$
- Attention vector
	- Query and key vectors between token $i$ and $j$
	- $a_{ij}=q_i\cdot k_j$
	- Divided by root of dimensionality of key vectors
		- $\sqrt{d_k}$
	- Pass through softmax to normalise
- $W_Q$ and $W_K$ are different matrices
	- Attention can be non-symmetric
	- Token $i$ attends to $j$ ($q_i\cdot k_j$ is large)
		- Doesn't imply that $j$ attends to $i$ ($q_j\cdot k_i$ can be small)
- Output for token $i$ is weighted sum of value vectors of all tokens weighted by $a_{ij}$
	- Attention from token $i$ to each other token
- $Q, K, V$ are matrices where $i$th row are vectors $q_i, k_i, v_i$ respectively 
$$\text{Attention}(Q,K,V)=\text{softmax}\left( \frac{QK^T}{\sqrt{d_k}} \right)V$$
- softmax taken over horizontal axis