Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention

Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, François Fleuret
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:5156-5165, 2020.

Abstract

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 \emph{Linear Transformers} achieve similar performance to vanilla Transformers and they are up to 4000x faster on autoregressive prediction of very long sequences.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-katharopoulos20a, title = {Transformers are {RNN}s: Fast Autoregressive Transformers with Linear Attention}, author = {Katharopoulos, Angelos and Vyas, Apoorv and Pappas, Nikolaos and Fleuret, Fran{\c{c}}ois}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {5156--5165}, year = {2020}, editor = {Hal Daumé III and Aarti Singh}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/katharopoulos20a/katharopoulos20a.pdf}, url = { http://proceedings.mlr.press/v119/katharopoulos20a.html }, abstract = {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 \emph{Linear Transformers} achieve similar performance to vanilla Transformers and they are up to 4000x faster on autoregressive prediction of very long sequences.} }
Endnote
%0 Conference Paper %T Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention %A Angelos Katharopoulos %A Apoorv Vyas %A Nikolaos Pappas %A François Fleuret %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-katharopoulos20a %I PMLR %P 5156--5165 %U http://proceedings.mlr.press/v119/katharopoulos20a.html %V 119 %X 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 \emph{Linear Transformers} achieve similar performance to vanilla Transformers and they are up to 4000x faster on autoregressive prediction of very long sequences.
APA
Katharopoulos, A., Vyas, A., Pappas, N. & Fleuret, F.. (2020). Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:5156-5165 Available from http://proceedings.mlr.press/v119/katharopoulos20a.html .

Related Material