Efficient Transformers

Linear Transformer

Architecture using kernelized decomposition of attention to achieve linear complexity in sequence and memory. Linear Transformer replaces softmax with positive kernel functions to enable associative reordering.

← Tillbaka