Cơ chế chú ý

The 2017 Attention Is All You Need paper introduced the transformer architecture built around attention rather than recurrence. Attention can be described through queries, keys, and values. A query is compared with keys to produce scores, and normalized weights combine the corresponding value vectors. Self-attention uses representations from the same sequence, while cross-attention can obtain keys and values from a separate encoded sequence. The allowed positions are important: a causal decoder mask prevents a position from using future positions during autoregressive training or generation. It is therefore inaccurate to say that every self-attention layer always reads every token without restrictions. Scaled dot-product attention divides query-key dot products by the square root of the key dimension before softmax. This scaling addresses how score magnitudes grow with dimension under the paper's assumptions. Softmax produces normalized weights over allowed positions, and the weighted value sum becomes the attention output. In a simple scalar illustration, values two and six with weights three-quarters and one-quarter combine to three. Multi-head attention performs several learned projections and attention operations before combining their outputs. It provides capacity to represent different relationships or subspaces without guaranteeing that every head has a unique, human-readable semantic role. Dense full attention involves pairwise interactions that scale quadratically with sequence length. Efficient implementations and alternative attention patterns address different costs; their benefits need measurement for a particular workload.