We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.CL

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo

Computer Science > Computation and Language

Title: Sparsifying Transformer Models with Trainable Representation Pooling

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 $1.8\times$ faster during training, $4.5\times$ faster during inference, and up to $13\times$ more computationally efficient in the decoder.
Comments: Accepted at ACL 2022
Subjects: Computation and Language (cs.CL); Machine Learning (cs.LG)
Cite as: arXiv:2009.05169 [cs.CL]
  (or arXiv:2009.05169v4 [cs.CL] for this version)

Submission history

From: Lukasz Borchmann [view email]
[v1] Thu, 10 Sep 2020 22:49:39 GMT (158kb,D)
[v2] Thu, 14 Jan 2021 16:49:01 GMT (263kb,D)
[v3] Thu, 4 Feb 2021 22:34:47 GMT (975kb,D)
[v4] Mon, 7 Mar 2022 12:49:08 GMT (1010kb,D)

Link back to: arXiv, form interface, contact.