Multi-Head Attention
Running several attention operations in parallel over different projections.
Overview
Multi-head attention splits the representation into several subspaces, runs attention independently in each, and concatenates the results. It solves the problem that a single attention distribution has to average over incompatible kinds of relationship - syntactic, semantic, positional - in one weighting. It exists because several narrow attentions outperform one wide one at the same cost.
Technology Graph
How Multi-Head Attention connects to the rest of the stack.
Classification
- Transformer architecture Algorithms Multi-Head Attention part of Transformer High ✓ source
Tools
Software that implements or integrates with Multi-Head Attention.