Architecture
aka MHA

Multi-head Attention

Attention applied in parallel across several projections, each capturing different relations.

Definition

Multi-head attention splits the model dimension into multiple smaller heads, each computing its own scaled dot-product attention. The results are concatenated and projected, letting the network attend to different relational patterns simultaneously — syntactic, semantic, positional.

Common use cases

  • Every transformer
  • Vision models
  • Speech models

Related terms

    Multi-head Attention — AI Glossary | Railwail