AI
layer norm
stabilizes activations
Source board
This definition comes from the board below. Open it to see the term in context.
Attention and the TransformerAIMore terms from this board
- attention matrixn×n who-attends-to-whom
- causal maskcan only attend to earlier tokens (for generation)
- cross-attentionattend to another sequence
- feed-forward (FFN/MLP)the per-token 'thinking' layer
- multi-headparallel attentions capturing different relations
- O(n²)attention's cost in sequence length (the context bottleneck)
- Query/Key/Value'what I seek / what I offer / what I give'
- residual connectionadd the input back (deep-net enabler)
- scaled dot-product attentionsoftmax(QKᵀ/√d)V
- self-attentioneach token attends to the others in its own sequence
- stack of N blocksthe depth of the model
- transformer blockattention + FFN + residual + LayerNorm