When can transformers reason with abstract symbols?

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Boix-Adsera, Enric, Saremi, Omid, Abbe, Emmanuel, Bengio, Samy, Littwin, Etai, Susskind, Joshua
Format: Preprint
Published: 2023
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866910411570282496
author Boix-Adsera, Enric
Saremi, Omid
Abbe, Emmanuel
Bengio, Samy
Littwin, Etai
Susskind, Joshua
author_facet Boix-Adsera, Enric
Saremi, Omid
Abbe, Emmanuel
Bengio, Samy
Littwin, Etai
Susskind, Joshua
contents We investigate the capabilities of transformer models on relational reasoning tasks. In these tasks, models are trained on a set of strings encoding abstract relations, and are then tested out-of-distribution on data that contains symbols that did not appear in the training dataset. We prove that for any relational reasoning task in a large family of tasks, transformers learn the abstract relations and generalize to the test set when trained by gradient descent on sufficiently large quantities of training data. This is in contrast to classical fully-connected networks, which we prove fail to learn to reason. Our results inspire modifications of the transformer architecture that add only two trainable parameters per head, and that we empirically demonstrate improve data efficiency for learning to reason.
format Preprint
id arxiv_https___arxiv_org_abs_2310_09753
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle When can transformers reason with abstract symbols?
Boix-Adsera, Enric
Saremi, Omid
Abbe, Emmanuel
Bengio, Samy
Littwin, Etai
Susskind, Joshua
Computation and Language
Artificial Intelligence
Machine Learning
We investigate the capabilities of transformer models on relational reasoning tasks. In these tasks, models are trained on a set of strings encoding abstract relations, and are then tested out-of-distribution on data that contains symbols that did not appear in the training dataset. We prove that for any relational reasoning task in a large family of tasks, transformers learn the abstract relations and generalize to the test set when trained by gradient descent on sufficiently large quantities of training data. This is in contrast to classical fully-connected networks, which we prove fail to learn to reason. Our results inspire modifications of the transformer architecture that add only two trainable parameters per head, and that we empirically demonstrate improve data efficiency for learning to reason.
title When can transformers reason with abstract symbols?
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2310.09753