Md. Asif Uddin

    Proposition 1II.3.P0143 of 86 in the corpus

    Attention is a weighted average whose weights are computed from content.

    Attention outputs a convex combination of value vectors, weighted by how well each key matches the query. The weights come from content; the position of a token plays no part.

    Attention as a weighted average over valuesA single query is compared against four keys. The resulting softmax weights — 0.06, 0.61, 0.09 and 0.24 — are shown as horizontal bars, and the output is the sum of the value vectors scaled by those weights. The weights come from content, not from position.query qsoftmax(q · kᵢ / √d)k₁0.06k₂0.61k₃0.09k₄0.24Σ wᵢ vᵢweights are content-addressed — nothing here depends on i
    Fig. 1 — Attention as a weighted average over values, with weights computed by comparing a query against every key. Nothing in the computation depends on position.

    Demonstration

    Attention is often introduced as a mechanism for “focus”, which is evocative and not very useful. Mechanically it is an average.

    For a query vector q and a set of key–value pairs (kᵢ, vᵢ):

    wᵢ = softmax( q · kᵢ / √d )
out = Σ wᵢ vᵢ

    The output is a convex combination of the value vectors. It lives in the convex hull of the values and can never leave it — which is why the feed-forward block after attention is not optional decoration. Attention mixes; the MLP transforms.

    The division by √d is not cosmetic. For q and k with unit-variance components, the dot product has variance d. Left unscaled, the logits grow with dimension, the softmax saturates, and gradients through it vanish. Dividing by √d restores unit variance and keeps the distribution soft enough to train.

    The decisive property is where the weights come from: content. The index i appears nowhere in the computation of wᵢ. Two identical keys receive identical weights no matter where in the sequence they sit.

    Corollary

    Because the weights are content-addressed and index-free, the whole operation is blind to order. That blindness is the subject of Chapter VI, and the whole of that chapter is the repair.

    Sources

    Depends on

    Used by

    Problems using this