الانتباه الذاتي: Query وKey وValue

In a sequence, the meaning of one word depends on others. Self-attention lets every token look at every other token to gather the context it needs.

Each token plays three roles: a query that asks, a key that answers, and a value that carries content. These come from the same word, used three ways.

A token's query describes what it is looking for. Think of it as the question this word is asking about the rest of the sentence.

Every token also exposes a key, a label advertising what it offers. A query is compared against all keys to find good matches.

Once a match is found, the value is the actual information that gets passed along. Keys decide how much, values decide what.

You build queries, keys, and values by projecting the input through three learned linear layers. Same input, three different weight matrices.

To see how well a query matches a key, you take their dot product. A bigger score means the two tokens are more relevant to each other.

Raw scores become attention weights with softmax, so each query's weights are positive and sum to one across all keys.

The output for each token is a weighted sum of all values, mixed by the attention weights. Relevant tokens contribute more.

Self-attention connects any two tokens in one step, so distance does not matter. This parallel view is why transformers handle long context so well.

The Q, K, V projections are trained by gradient descent. The network learns what to ask, what to advertise, and what to share, all from data.

You learned that self-attention turns each token into a query, key, and value, scores query-key matches, and blends values by softmax weights. Nice work!