Attention Mechanism

The attention mechanism is the core innovation of the transformer architecture. It allows a model to compute a weighted representation of all positions in an input sequence when processing any given position, with higher weights assigned to positions that are more relevant to the current computation. This enables the model to capture long-range dependencies without processing sequences step-by-step.

Formally, attention computes three quantities for each position: queries, keys, and values. The query represents what the current position is looking for. The keys represent what each position offers. The attention weight for a given (query, key) pair is computed as the dot product of the query and key, scaled and normalized through a softmax function. The output is then a weighted sum of the values, where the weights are the attention scores.

Multi-head attention runs multiple independent attention operations in parallel (the 'heads'), each learning to attend to different types of relationships. One head might learn to track syntactic subject-verb agreement while another captures coreference relationships and another attends to topical similarity. The outputs of all heads are concatenated and linearly projected to produce the final attention output.

Self-attention, used in the transformer's encoder and decoder layers, allows each position in the sequence to attend to every other position. This is what gives transformer models their ability to relate distant parts of a document: a pronoun can attend directly to the noun it refers to, regardless of how many words separate them.

The attention mechanism is why the quality and ordering of information in a prompt affects model outputs so substantially.