BoRA: Bi-dimensional Weight-Decomposed Low-Rank Adaptation

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Wang, Qiushi, Fan, Yuchen, Bao, Junwei, Jiang, Hongfei, Song, Yang
Natura: Preprint
Pubblicazione: 2024
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866929620612284416
author Wang, Qiushi
Fan, Yuchen
Bao, Junwei
Jiang, Hongfei
Song, Yang
author_facet Wang, Qiushi
Fan, Yuchen
Bao, Junwei
Jiang, Hongfei
Song, Yang
contents In recent years, Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) have significantly enhanced the adaptability of large-scale pre-trained models. Weight-Decomposed Low-Rank Adaptation (DoRA) improves upon LoRA by separating the magnitude and direction components of the weight matrix, leading to superior performance. However, DoRA's improvements are limited to the vertical dimension, resulting in an asymmetrical pattern between horizontal and vertical dimensions. This paper introduces BoRA, an innovative extension of LoRA and DoRA, characterized by symmetrical properties across horizontal and vertical dimensions. Our approach optimizes the weight matrix symmetrically by adjusting both column-wise and row-wise magnitudes. Extensive experiments demonstrate that BoRA surpasses state-of-the-art PEFT methods, including LoRA and DoRA, achieving superior results across various benchmarks.
format Preprint
id arxiv_https___arxiv_org_abs_2412_06441
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle BoRA: Bi-dimensional Weight-Decomposed Low-Rank Adaptation
Wang, Qiushi
Fan, Yuchen
Bao, Junwei
Jiang, Hongfei
Song, Yang
Computation and Language
In recent years, Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) have significantly enhanced the adaptability of large-scale pre-trained models. Weight-Decomposed Low-Rank Adaptation (DoRA) improves upon LoRA by separating the magnitude and direction components of the weight matrix, leading to superior performance. However, DoRA's improvements are limited to the vertical dimension, resulting in an asymmetrical pattern between horizontal and vertical dimensions. This paper introduces BoRA, an innovative extension of LoRA and DoRA, characterized by symmetrical properties across horizontal and vertical dimensions. Our approach optimizes the weight matrix symmetrically by adjusting both column-wise and row-wise magnitudes. Extensive experiments demonstrate that BoRA surpasses state-of-the-art PEFT methods, including LoRA and DoRA, achieving superior results across various benchmarks.
title BoRA: Bi-dimensional Weight-Decomposed Low-Rank Adaptation
topic Computation and Language
url https://arxiv.org/abs/2412.06441