Sparsifying Transformer Models with Trainable Representation Pooling

Michał Pietruszka, Łukasz Borchmann, Łukasz Garncarek


Abstract
We propose a novel method to sparsify attention in the Transformer model by learning to select the most-informative token representations during the training process, thus focusing on the task-specific parts of an input. A reduction of quadratic time and memory complexity to sublinear was achieved due to a robust trainable top-k operator.Our experiments on a challenging long document summarization task show that even our simple baseline performs comparably to the current SOTA, and with trainable pooling we can retain its top quality, while being 1.8× faster during training, 4.5× faster during inference, and up to 13× more computationally efficient in the decoder.
Anthology ID:
2022.acl-long.590
Volume:
Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)
Month:
May
Year:
2022
Address:
Dublin, Ireland
Venue:
ACL
SIG:
Publisher:
Association for Computational Linguistics
Note:
Pages:
8616–8633
Language:
URL:
https://aclanthology.org/2022.acl-long.590
DOI:
10.18653/v1/2022.acl-long.590
Bibkey:
Cite (ACL):
Michał Pietruszka, Łukasz Borchmann, and Łukasz Garncarek. 2022. Sparsifying Transformer Models with Trainable Representation Pooling. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8616–8633, Dublin, Ireland. Association for Computational Linguistics.
Cite (Informal):
Sparsifying Transformer Models with Trainable Representation Pooling (Pietruszka et al., ACL 2022)
Copy Citation:
PDF:
https://aclanthology.org/2022.acl-long.590.pdf
Software:
 2022.acl-long.590.software.zip
Code
 applicaai/pyramidions
Data
PubmedarXiv Summarization Dataset