Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Kajitsuka, Tokio, Sato, Issei
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911765800943616
author Kajitsuka, Tokio
Sato, Issei
author_facet Kajitsuka, Tokio
Sato, Issei
contents Existing analyses of the expressive capacity of Transformer models have required excessively deep layers for data memorization, leading to a discrepancy with the Transformers actually used in practice. This is primarily due to the interpretation of the softmax function as an approximation of the hardmax function. By clarifying the connection between the softmax function and the Boltzmann operator, we prove that a single layer of self-attention with low-rank weight matrices possesses the capability to perfectly capture the context of an entire input sequence. As a consequence, we show that one-layer and single-head Transformers have a memorization capacity for finite samples, and that Transformers consisting of one self-attention layer with two feed-forward neural networks are universal approximators for continuous permutation equivariant functions on a compact domain.
format Preprint
id arxiv_https___arxiv_org_abs_2307_14023
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators?
Kajitsuka, Tokio
Sato, Issei
Machine Learning
68T07
I.2.0
Existing analyses of the expressive capacity of Transformer models have required excessively deep layers for data memorization, leading to a discrepancy with the Transformers actually used in practice. This is primarily due to the interpretation of the softmax function as an approximation of the hardmax function. By clarifying the connection between the softmax function and the Boltzmann operator, we prove that a single layer of self-attention with low-rank weight matrices possesses the capability to perfectly capture the context of an entire input sequence. As a consequence, we show that one-layer and single-head Transformers have a memorization capacity for finite samples, and that Transformers consisting of one self-attention layer with two feed-forward neural networks are universal approximators for continuous permutation equivariant functions on a compact domain.
title Are Transformers with One Layer Self-Attention Using Low-Rank Weight Matrices Universal Approximators?
topic Machine Learning
68T07
I.2.0
url https://arxiv.org/abs/2307.14023