Decoupling the Role of Data, Attention, and Losses in Multimodal Transformers

Lisa Anne Hendricks, John Mellor, Rosalia Schneider, Jean-Baptiste Alayrac, Aida Nematzadeh


Abstract
Recently, multimodal transformer models have gained popularity because their performance on downstream tasks suggests they learn rich visual-linguistic representations. Focusing on zero-shot image retrieval tasks, we study three important factors that can impact the quality of learned representations: pretraining data, the attention mechanism, and loss functions. By pretraining models on six datasets, we observe that dataset noise and language similarity to our downstream task are important indicators of model performance. Through architectural analysis, we learn that models with a multimodal attention mechanism can outperform deeper models with modality-specific attention mechanisms. Finally, we show that successful contrastive losses used in the self-supervised learning literature do not yield similar performance gains when used in multimodal transformers.
Anthology ID:
2021.tacl-1.35
Volume:
Transactions of the Association for Computational Linguistics, Volume 9
Month:
Year:
2021
Address:
Cambridge, MA
Editors:
Brian Roark, Ani Nenkova
Venue:
TACL
SIG:
Publisher:
MIT Press
Note:
Pages:
570–585
Language:
URL:
https://aclanthology.org/2021.tacl-1.35
DOI:
10.1162/tacl_a_00385
Bibkey:
Cite (ACL):
Lisa Anne Hendricks, John Mellor, Rosalia Schneider, Jean-Baptiste Alayrac, and Aida Nematzadeh. 2021. Decoupling the Role of Data, Attention, and Losses in Multimodal Transformers. Transactions of the Association for Computational Linguistics, 9:570–585.
Cite (Informal):
Decoupling the Role of Data, Attention, and Losses in Multimodal Transformers (Hendricks et al., TACL 2021)
Copy Citation:
PDF:
https://preview.aclanthology.org/improve-issue-templates/2021.tacl-1.35.pdf
Video:
 https://preview.aclanthology.org/improve-issue-templates/2021.tacl-1.35.mp4