The content discusses the challenges of multi-task model fusion with large pre-trained models and introduces a novel partial linearization method to improve fusion efficiency. The proposed approach demonstrates superior performance in constructing unified multi-task models compared to standard methods.
Large pre-trained models have enabled significant advances in machine learning, but efficiently fine-tuning them on multiple tasks remains challenging. The proposed partial linearization technique improves multi-task fusion by leveraging model fusion benefits over linearized fine-tuning. Experimental results show enhanced performance across various tasks, highlighting the effectiveness of the approach.
Parameter-efficient fine-tuning techniques reduce computational costs but can lead to representational interference between tasks. The key challenge is performing PEFT while preventing negative interference between task-specific representations. The proposed method aims to enhance multi-task model fusion capabilities for parameter-efficient fine-tuned models.
Model fusion methods extract knowledge from models fine-tuned on different tasks to create a unified model that performs well across multiple tasks. The content explores various fusion algorithms and compares their performance using different fine-tuning methods like LoRA and L-LoRA.
To Another Language
from source content
arxiv.org
Viktige innsikter hentet fra
by Anke Tang,Li... klokken arxiv.org 03-12-2024
https://arxiv.org/pdf/2310.04742.pdfDypere Spørsmål