Pay Less Attention with Lightweight and Dynamic Convolutions. Wu, F., Fan, A., Baevski, A., Dauphin, Y., N., & Auli, M. 7th International Conference on Learning Representations, ICLR 2019, International Conference on Learning Representations, ICLR, 1, 2019.
Pay Less Attention with Lightweight and Dynamic Convolutions [link]Website  abstract   bibtex   
Self-attention is a useful mechanism to build generative models for language and images. It determines the importance of context elements by comparing each element to the current time step. In this paper, we show that a very lightweight convolution can perform competitively to the best reported self-attention results. Next, we introduce dynamic convolutions which are simpler and more efficient than self-attention. We predict separate convolution kernels based solely on the current time-step in order to determine the importance of context elements. The number of operations required by this approach scales linearly in the input length, whereas self-attention is quadratic. Experiments on large-scale machine translation, language modeling and abstractive summarization show that dynamic convolutions improve over strong self-attention models. On the WMT'14 English-German test set dynamic convolutions achieve a new state of the art of 29.7 BLEU.
@article{
 title = {Pay Less Attention with Lightweight and Dynamic Convolutions},
 type = {article},
 year = {2019},
 websites = {https://arxiv.org/abs/1901.10430v1},
 month = {1},
 publisher = {International Conference on Learning Representations, ICLR},
 day = {29},
 id = {a4bf96df-3972-3780-baf4-0bce55c21b88},
 created = {2021-08-24T07:15:01.585Z},
 accessed = {2021-08-24},
 file_attached = {true},
 profile_id = {48fc0258-023d-3602-860e-824092d62c56},
 group_id = {1ff583c0-be37-34fa-9c04-73c69437d354},
 last_modified = {2021-08-24T07:15:04.707Z},
 read = {false},
 starred = {false},
 authored = {false},
 confirmed = {false},
 hidden = {false},
 folder_uuids = {c509f25c-b687-4ab5-8859-72131b6658d3},
 private_publication = {false},
 abstract = {Self-attention is a useful mechanism to build generative models for language
and images. It determines the importance of context elements by comparing each
element to the current time step. In this paper, we show that a very
lightweight convolution can perform competitively to the best reported
self-attention results. Next, we introduce dynamic convolutions which are
simpler and more efficient than self-attention. We predict separate convolution
kernels based solely on the current time-step in order to determine the
importance of context elements. The number of operations required by this
approach scales linearly in the input length, whereas self-attention is
quadratic. Experiments on large-scale machine translation, language modeling
and abstractive summarization show that dynamic convolutions improve over
strong self-attention models. On the WMT'14 English-German test set dynamic
convolutions achieve a new state of the art of 29.7 BLEU.},
 bibtype = {article},
 author = {Wu, Felix and Fan, Angela and Baevski, Alexei and Dauphin, Yann N. and Auli, Michael},
 journal = {7th International Conference on Learning Representations, ICLR 2019}
}

Downloads: 0