The Expressive Capacity of State Space Models: A Formal Language Perspective

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Sarrof, Yash, Veitsman, Yana, Hahn, Michael
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912758630449152
author Sarrof, Yash
Veitsman, Yana
Hahn, Michael
author_facet Sarrof, Yash
Veitsman, Yana
Hahn, Michael
contents Recently, recurrent models based on linear state space models (SSMs) have shown promising performance in language modeling (LM), competititve with transformers. However, there is little understanding of the in-principle abilities of such models, which could provide useful guidance to the search for better LM architectures. We present a comprehensive theoretical study of the capacity of such SSMs as it compares to that of transformers and traditional RNNs. We find that SSMs and transformers have overlapping but distinct strengths. In star-free state tracking, SSMs implement straightforward and exact solutions to problems that transformers struggle to represent exactly. They can also model bounded hierarchical structure with optimal memory even without simulating a stack. On the other hand, we identify a design choice in current SSMs that limits their expressive power. We discuss implications for SSM and LM research, and verify results empirically on a recent SSM, Mamba.
format Preprint
id arxiv_https___arxiv_org_abs_2405_17394
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle The Expressive Capacity of State Space Models: A Formal Language Perspective
Sarrof, Yash
Veitsman, Yana
Hahn, Michael
Computation and Language
Formal Languages and Automata Theory
Machine Learning
Recently, recurrent models based on linear state space models (SSMs) have shown promising performance in language modeling (LM), competititve with transformers. However, there is little understanding of the in-principle abilities of such models, which could provide useful guidance to the search for better LM architectures. We present a comprehensive theoretical study of the capacity of such SSMs as it compares to that of transformers and traditional RNNs. We find that SSMs and transformers have overlapping but distinct strengths. In star-free state tracking, SSMs implement straightforward and exact solutions to problems that transformers struggle to represent exactly. They can also model bounded hierarchical structure with optimal memory even without simulating a stack. On the other hand, we identify a design choice in current SSMs that limits their expressive power. We discuss implications for SSM and LM research, and verify results empirically on a recent SSM, Mamba.
title The Expressive Capacity of State Space Models: A Formal Language Perspective
topic Computation and Language
Formal Languages and Automata Theory
Machine Learning
url https://arxiv.org/abs/2405.17394