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

Download:

Current browse context:

cs.LG

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 ScienceWISE logo

Computer Science > Machine Learning

Title: Are Transformers universal approximators of sequence-to-sequence functions?

Abstract: Despite the widespread adoption of Transformer models for NLP tasks, the expressive power of these models is not well-understood. In this paper, we establish that Transformer models are universal approximators of continuous permutation equivariant sequence-to-sequence functions with compact support, which is quite surprising given the amount of shared parameters in these models. Furthermore, using positional encodings, we circumvent the restriction of permutation equivariance, and show that Transformer models can universally approximate arbitrary continuous sequence-to-sequence functions on a compact domain. Interestingly, our proof techniques clearly highlight the different roles of the self-attention and the feed-forward layers in Transformers. In particular, we prove that fixed width self-attention layers can compute contextual mappings of the input sequences, playing a key role in the universal approximation property of Transformers. Based on this insight from our analysis, we consider other simpler alternatives to self-attention layers and empirically evaluate them.
Comments: 23 pages, ICLR 2020 camera-ready version
Subjects: Machine Learning (cs.LG); Machine Learning (stat.ML)
Cite as: arXiv:1912.10077 [cs.LG]
  (or arXiv:1912.10077v2 [cs.LG] for this version)

Submission history

From: Chulhee Yun [view email]
[v1] Fri, 20 Dec 2019 19:49:32 GMT (196kb,D)
[v2] Tue, 25 Feb 2020 03:12:57 GMT (199kb,D)

Link back to: arXiv, form interface, contact.