Multi-Head Attention

Linearized Attention

Family of attention mechanisms that rewrite the attention calculation to avoid materializing the full attention matrix, allowing linear complexity relative to the sequence length.

← Indietro