Magneto: A Foundation Transformer

Hongyu Wang, Shuming Ma, Shaohan Huang, Li Dong, Wenhui Wang, Zhiliang Peng, Yu Wu, Payal Bajaj, Saksham Singhal, Alon Benhaim, Barun Patra, Zhun Liu, Vishrav Chaudhary, Xia Song, Furu Wei
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:36077-36092, 2023.

Abstract

A big convergence of model architectures across language, vision, speech, and multimodal is emerging. However, under the same name ”Transformers”, the above areas use different implementations for better performance, e.g., Post-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We call for the development of Foundation Transformer for true general-purpose modeling, which serves as a go-to architecture for various tasks and modalities with guaranteed training stability. In this work, we introduce a Transformer variant, named Magneto, to fulfill the goal. Specifically, we propose Sub-LayerNorm for good expressivity, and the initialization strategy theoretically derived from DeepNet for stable scaling up. Extensive experiments demonstrate its superior performance and better stability than the de facto Transformer variants designed for various applications, including language modeling (i.e., BERT, and GPT), machine translation, vision pretraining (i.e., BEiT), speech recognition, and multimodal pretraining (i.e., BEiT-3).

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-wang23u, title = {Magneto: A Foundation Transformer}, author = {Wang, Hongyu and Ma, Shuming and Huang, Shaohan and Dong, Li and Wang, Wenhui and Peng, Zhiliang and Wu, Yu and Bajaj, Payal and Singhal, Saksham and Benhaim, Alon and Patra, Barun and Liu, Zhun and Chaudhary, Vishrav and Song, Xia and Wei, Furu}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {36077--36092}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/wang23u/wang23u.pdf}, url = {https://proceedings.mlr.press/v202/wang23u.html}, abstract = {A big convergence of model architectures across language, vision, speech, and multimodal is emerging. However, under the same name ”Transformers”, the above areas use different implementations for better performance, e.g., Post-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We call for the development of Foundation Transformer for true general-purpose modeling, which serves as a go-to architecture for various tasks and modalities with guaranteed training stability. In this work, we introduce a Transformer variant, named Magneto, to fulfill the goal. Specifically, we propose Sub-LayerNorm for good expressivity, and the initialization strategy theoretically derived from DeepNet for stable scaling up. Extensive experiments demonstrate its superior performance and better stability than the de facto Transformer variants designed for various applications, including language modeling (i.e., BERT, and GPT), machine translation, vision pretraining (i.e., BEiT), speech recognition, and multimodal pretraining (i.e., BEiT-3).} }
Endnote
%0 Conference Paper %T Magneto: A Foundation Transformer %A Hongyu Wang %A Shuming Ma %A Shaohan Huang %A Li Dong %A Wenhui Wang %A Zhiliang Peng %A Yu Wu %A Payal Bajaj %A Saksham Singhal %A Alon Benhaim %A Barun Patra %A Zhun Liu %A Vishrav Chaudhary %A Xia Song %A Furu Wei %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-wang23u %I PMLR %P 36077--36092 %U https://proceedings.mlr.press/v202/wang23u.html %V 202 %X A big convergence of model architectures across language, vision, speech, and multimodal is emerging. However, under the same name ”Transformers”, the above areas use different implementations for better performance, e.g., Post-LayerNorm for BERT, and Pre-LayerNorm for GPT and vision Transformers. We call for the development of Foundation Transformer for true general-purpose modeling, which serves as a go-to architecture for various tasks and modalities with guaranteed training stability. In this work, we introduce a Transformer variant, named Magneto, to fulfill the goal. Specifically, we propose Sub-LayerNorm for good expressivity, and the initialization strategy theoretically derived from DeepNet for stable scaling up. Extensive experiments demonstrate its superior performance and better stability than the de facto Transformer variants designed for various applications, including language modeling (i.e., BERT, and GPT), machine translation, vision pretraining (i.e., BEiT), speech recognition, and multimodal pretraining (i.e., BEiT-3).
APA
Wang, H., Ma, S., Huang, S., Dong, L., Wang, W., Peng, Z., Wu, Y., Bajaj, P., Singhal, S., Benhaim, A., Patra, B., Liu, Z., Chaudhary, V., Song, X. & Wei, F.. (2023). Magneto: A Foundation Transformer. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:36077-36092 Available from https://proceedings.mlr.press/v202/wang23u.html.

Related Material