Improving Transformers with Dynamically Composable Multi-Head Attention

Da Xiao, Qingye Meng, Shengping Li, Xingyuan Yuan
Proceedings of the 41st International Conference on Machine Learning, PMLR 235:54300-54318, 2024.

Abstract

Multi-Head Attention (MHA) is a key component of Transformer. In MHA, attention heads work independently, causing problems such as low-rank bottleneck of attention score matrices and head redundancy. We propose Dynamically Composable Multi-Head Attention (DCMHA), a parameter and computation efficient attention architecture that tackles the shortcomings of MHA and increases the expressive power of the model by dynamically composing attention heads. At the core of DCMHA is a Compose function that transforms the attention score and weight matrices in an input-dependent way. DCMHA can be used as a drop-in replacement of MHA in any transformer architecture to obtain the corresponding DCFormer. DCFormer significantly outperforms Transformer on different architectures and model scales in language modeling, matching the performance of models with 1.7x-2.0x compute. For example, DCPythia-6.9B outperforms open source Pythia-12B on both pretraining perplexity and downstream task evaluation.

Cite this Paper

BibTeX
@InProceedings{pmlr-v235-xiao24d, title = {Improving Transformers with Dynamically Composable Multi-Head Attention}, author = {Xiao, Da and Meng, Qingye and Li, Shengping and Yuan, Xingyuan}, booktitle = {Proceedings of the 41st International Conference on Machine Learning}, pages = {54300--54318}, year = {2024}, editor = {Salakhutdinov, Ruslan and Kolter, Zico and Heller, Katherine and Weller, Adrian and Oliver, Nuria and Scarlett, Jonathan and Berkenkamp, Felix}, volume = {235}, series = {Proceedings of Machine Learning Research}, month = {21--27 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v235/main/assets/xiao24d/xiao24d.pdf}, url = {https://proceedings.mlr.press/v235/xiao24d.html}, abstract = {Multi-Head Attention (MHA) is a key component of Transformer. In MHA, attention heads work independently, causing problems such as low-rank bottleneck of attention score matrices and head redundancy. We propose Dynamically Composable Multi-Head Attention (DCMHA), a parameter and computation efficient attention architecture that tackles the shortcomings of MHA and increases the expressive power of the model by dynamically composing attention heads. At the core of DCMHA is a Compose function that transforms the attention score and weight matrices in an input-dependent way. DCMHA can be used as a drop-in replacement of MHA in any transformer architecture to obtain the corresponding DCFormer. DCFormer significantly outperforms Transformer on different architectures and model scales in language modeling, matching the performance of models with 1.7x-2.0x compute. For example, DCPythia-6.9B outperforms open source Pythia-12B on both pretraining perplexity and downstream task evaluation.} }
Endnote
%0 Conference Paper %T Improving Transformers with Dynamically Composable Multi-Head Attention %A Da Xiao %A Qingye Meng %A Shengping Li %A Xingyuan Yuan %B Proceedings of the 41st International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2024 %E Ruslan Salakhutdinov %E Zico Kolter %E Katherine Heller %E Adrian Weller %E Nuria Oliver %E Jonathan Scarlett %E Felix Berkenkamp %F pmlr-v235-xiao24d %I PMLR %P 54300--54318 %U https://proceedings.mlr.press/v235/xiao24d.html %V 235 %X Multi-Head Attention (MHA) is a key component of Transformer. In MHA, attention heads work independently, causing problems such as low-rank bottleneck of attention score matrices and head redundancy. We propose Dynamically Composable Multi-Head Attention (DCMHA), a parameter and computation efficient attention architecture that tackles the shortcomings of MHA and increases the expressive power of the model by dynamically composing attention heads. At the core of DCMHA is a Compose function that transforms the attention score and weight matrices in an input-dependent way. DCMHA can be used as a drop-in replacement of MHA in any transformer architecture to obtain the corresponding DCFormer. DCFormer significantly outperforms Transformer on different architectures and model scales in language modeling, matching the performance of models with 1.7x-2.0x compute. For example, DCPythia-6.9B outperforms open source Pythia-12B on both pretraining perplexity and downstream task evaluation.
APA
Xiao, D., Meng, Q., Li, S. & Yuan, X.. (2024). Improving Transformers with Dynamically Composable Multi-Head Attention. Proceedings of the 41st International Conference on Machine Learning, in Proceedings of Machine Learning Research 235:54300-54318 Available from https://proceedings.mlr.press/v235/xiao24d.html.

Related Material