One head of a ferry-schedule model tracks who-did-what, another tracks times, and another tracks places. Why use several attention heads instead of one averaged view?
Select an answer to reveal the explanation.
Short Explanation
One averaged view is like asking one crew member to watch who, when, and where at the same time. You get a blur. Multi-head attention gives each head its own specialty so those relationship types stay separate. Extra heads do not drop embeddings or turn into cooling fans.
Full Explanation
Multiple heads can specialize (who, when, where) instead of collapsing every relationship into one average. They do not drop those relations, replace embeddings, or describe cooling zones.