Sparse Attention

Kernel-based Attention

Approach using kernels to approximate softmax attention, enabling linear complexity calculations through techniques like FAVOR+ (Fast Attention Via Positive Orthogonal Random Features).

← Tillbaka