Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Transformers are RNNs: Fast Autoregressive Transformers with Linear Attention

About

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 $\mathcal{O}\left(N^2\right)$ to $\mathcal{O}\left(N\right)$, 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 linear transformers achieve similar performance to vanilla transformers and they are up to 4000x faster on autoregressive prediction of very long sequences.

Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, Fran\c{c}ois Fleuret• 2020

Related benchmarks

TaskDatasetResultRank
Commonsense ReasoningHellaSwag--
1896
Commonsense ReasoningWinoGrande--
1442
Language ModelingWikiText
PPL45.43
740
Language ModelingWikiText-103 (test)
Perplexity22.2
703
Natural Language UnderstandingGLUE
SST-284.63
551
Natural Language UnderstandingGLUE (dev)
SST-2 (Acc)91.51
529
Machine TranslationWMT En-De 2014 (test)
BLEU28.4
379
Language ModelingWikiText-103 (val)
PPL27.44
261
Image ClassificationCIFAR-10
Accuracy60
246
Commonsense ReasoningARC Challenge
Accuracy21.76
243
Showing 10 of 127 rows
...

Other info

Code

Follow for update