Scaling Transformer to 1M tokens and beyond with RMT

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Bulatov, Aydar, Kuratov, Yuri, Kapushev, Yermek, Burtsev, Mikhail S.
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866929234870534144
author Bulatov, Aydar
Kuratov, Yuri
Kapushev, Yermek
Burtsev, Mikhail S.
author_facet Bulatov, Aydar
Kuratov, Yuri
Kapushev, Yermek
Burtsev, Mikhail S.
contents A major limitation for the broader scope of problems solvable by transformers is the quadratic scaling of computational complexity with input size. In this study, we investigate the recurrent memory augmentation of pre-trained transformer models to extend input context length while linearly scaling compute. Our approach demonstrates the capability to store information in memory for sequences of up to an unprecedented two million tokens while maintaining high retrieval accuracy. Experiments with language modeling tasks show perplexity improvement as the number of processed input segments increases. These results underscore the effectiveness of our method, which has significant potential to enhance long-term dependency handling in natural language understanding and generation tasks, as well as enable large-scale context processing for memory-intensive applications.
format Preprint
id arxiv_https___arxiv_org_abs_2304_11062
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Scaling Transformer to 1M tokens and beyond with RMT
Bulatov, Aydar
Kuratov, Yuri
Kapushev, Yermek
Burtsev, Mikhail S.
Computation and Language
Artificial Intelligence
Machine Learning
A major limitation for the broader scope of problems solvable by transformers is the quadratic scaling of computational complexity with input size. In this study, we investigate the recurrent memory augmentation of pre-trained transformer models to extend input context length while linearly scaling compute. Our approach demonstrates the capability to store information in memory for sequences of up to an unprecedented two million tokens while maintaining high retrieval accuracy. Experiments with language modeling tasks show perplexity improvement as the number of processed input segments increases. These results underscore the effectiveness of our method, which has significant potential to enhance long-term dependency handling in natural language understanding and generation tasks, as well as enable large-scale context processing for memory-intensive applications.
title Scaling Transformer to 1M tokens and beyond with RMT
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2304.11062