Dan Luo , Kangfeng Zheng , Chunhua Wu , Xiujuan Wang , Jvjie Wang
{"title":"ERAT-DLoRA: Parameter-efficient tuning with enhanced range adaptation in time and depth aware dynamic LoRA","authors":"Dan Luo , Kangfeng Zheng , Chunhua Wu , Xiujuan Wang , Jvjie Wang","doi":"10.1016/j.neucom.2024.128778","DOIUrl":null,"url":null,"abstract":"<div><div>Despite their potential, the industrial deployment of large language models (LLMs) is constrained by traditional fine-tuning procedures that are both resource-intensive and time-consuming. Low-Rank Adaptation (LoRA) has emerged as a pioneering methodology for addressing these challenges. By integrating low-rank decomposition matrices into network weights to reduce trainable parameters, LoRA effectively accelerates the adaptation process. While research on LoRA primarily focuses on adjusting low-rank matrices, DyLoRA optimizes the rank-setting mechanism to avoid extensive effort in rank size training and searching. However, DyLoRA rank configuration mechanism has its own limitation. First, DyLoRA sets the same rank size for all the low-rank adaptation layers at each time step. Given that layers with different depth contain distinct information, they should have varying rank values to accurately capture their unique characteristics. Second, the truncated phase selected for ordering representation based on nested dropout regulation is only half dynamic, continuously dropping tail units, thereby limiting its ability to access information. In this work, we propose a novel technique, enhanced range adaptation in time and depth aware dynamic LoRA (ERAT-DLoRA) to address these problems. The ERAT-DLoRA method introduces a dynamic range to the truncated phase that makes the truncated phase fully dynamic. Additionally, we design a time and layer-aware dynamic rank to ensure appropriate adjustments at different time steps and layer levels. We evaluate our solution on natural languages understanding and language generation tasks. Extensive evaluation results demonstrate the effectiveness of the proposed method.</div></div>","PeriodicalId":19268,"journal":{"name":"Neurocomputing","volume":"614 ","pages":"Article 128778"},"PeriodicalIF":5.5000,"publicationDate":"2024-10-28","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Neurocomputing","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S0925231224015492","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q1","JCRName":"COMPUTER SCIENCE, ARTIFICIAL INTELLIGENCE","Score":null,"Total":0}
引用次数: 0
Abstract
Despite their potential, the industrial deployment of large language models (LLMs) is constrained by traditional fine-tuning procedures that are both resource-intensive and time-consuming. Low-Rank Adaptation (LoRA) has emerged as a pioneering methodology for addressing these challenges. By integrating low-rank decomposition matrices into network weights to reduce trainable parameters, LoRA effectively accelerates the adaptation process. While research on LoRA primarily focuses on adjusting low-rank matrices, DyLoRA optimizes the rank-setting mechanism to avoid extensive effort in rank size training and searching. However, DyLoRA rank configuration mechanism has its own limitation. First, DyLoRA sets the same rank size for all the low-rank adaptation layers at each time step. Given that layers with different depth contain distinct information, they should have varying rank values to accurately capture their unique characteristics. Second, the truncated phase selected for ordering representation based on nested dropout regulation is only half dynamic, continuously dropping tail units, thereby limiting its ability to access information. In this work, we propose a novel technique, enhanced range adaptation in time and depth aware dynamic LoRA (ERAT-DLoRA) to address these problems. The ERAT-DLoRA method introduces a dynamic range to the truncated phase that makes the truncated phase fully dynamic. Additionally, we design a time and layer-aware dynamic rank to ensure appropriate adjustments at different time steps and layer levels. We evaluate our solution on natural languages understanding and language generation tasks. Extensive evaluation results demonstrate the effectiveness of the proposed method.
期刊介绍:
Neurocomputing publishes articles describing recent fundamental contributions in the field of neurocomputing. Neurocomputing theory, practice and applications are the essential topics being covered.