The Hidden Attention of Mamba Models

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ali, Ameen, Zimerman, Itamar, Wolf, Lior
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917626814398464
author Ali, Ameen
Zimerman, Itamar
Wolf, Lior
author_facet Ali, Ameen
Zimerman, Itamar
Wolf, Lior
contents The Mamba layer offers an efficient selective state space model (SSM) that is highly effective in modeling multiple domains, including NLP, long-range sequence processing, and computer vision. Selective SSMs are viewed as dual models, in which one trains in parallel on the entire sequence via an IO-aware parallel scan, and deploys in an autoregressive manner. We add a third view and show that such models can be viewed as attention-driven models. This new perspective enables us to empirically and theoretically compare the underlying mechanisms to that of the self-attention layers in transformers and allows us to peer inside the inner workings of the Mamba model with explainability methods. Our code is publicly available.
format Preprint
id arxiv_https___arxiv_org_abs_2403_01590
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle The Hidden Attention of Mamba Models
Ali, Ameen
Zimerman, Itamar
Wolf, Lior
Machine Learning
F.2.2, I.2.7
F.2.2; I.2.7
The Mamba layer offers an efficient selective state space model (SSM) that is highly effective in modeling multiple domains, including NLP, long-range sequence processing, and computer vision. Selective SSMs are viewed as dual models, in which one trains in parallel on the entire sequence via an IO-aware parallel scan, and deploys in an autoregressive manner. We add a third view and show that such models can be viewed as attention-driven models. This new perspective enables us to empirically and theoretically compare the underlying mechanisms to that of the self-attention layers in transformers and allows us to peer inside the inner workings of the Mamba model with explainability methods. Our code is publicly available.
title The Hidden Attention of Mamba Models
topic Machine Learning
F.2.2, I.2.7
F.2.2; I.2.7
url https://arxiv.org/abs/2403.01590