Timezone: »

Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention
Angelos Katharopoulos · Apoorv Vyas · Nikolaos Pappas · François Fleuret

Wed Jul 15 01:00 PM -- 01:45 PM & Thu Jul 16 01:00 AM -- 01:45 AM (PDT) @
Transformers achieve remarkable performance in several tasks but due to their quadratic complexity, with respect to the input's length, they are prohibitively slow for very long sequences. To address this limitation, we express the self-attention as a linear dot-product of kernel feature maps and make use of the associativity property of matrix products to reduce the complexity from $\bigO{N^2}$ to $\bigO{N}$, where $N$ is the sequence length. We show that this formulation permits an iterative implementation that dramatically accelerates autoregressive transformers and reveals their relationship to recurrent neural networks. Our \textit{Linear Transformers} achieve similar performance to vanilla Transformers and they are up to 4000x faster on autoregressive prediction of very long sequences.

Author Information

Angelos Katharopoulos (Idiap & EPFL)
Apoorv Vyas (Idiap Research Institute and EPFL)
Nikolaos Pappas (University of Washington)
François Fleuret (University of Geneva)

More from the Same Authors