Query, Key, Value: A Differentiable Dictionary Lookup

Take the sentence again:

"The animal didn't cross the street because it was too tired."

The token "it" carries, by itself, almost no useful information.

Somewhere else — in "animal", maybe in "street", in "tired".

So "it" must look up information stored in other positions.

Now compare with programming you may already know: a dictionary / hash map.

• You look up a key → the map returns one payload.
• Exact match, discrete, instant.

An attention layer wants the same shape of operation — ask, retrieve, receive content — but with two twists:

1. The lookup must survive approximate matches (no exact string; only vectors).
2. The lookup must be differentiable, so the model can learn better lookups from the loss.

A hash map fails both.

The 2017 design solves both with one idea:

That single sentence is this whole topic. Everything else spells out the three players — query, key, value — and why each needs its own learned matrix.