Parameter-Efficient Fine-Tuning via Circular Convolution

TL;DR

This paper introduces Circular Convolution Adaptation (C3A), a novel method for fine-tuning large models that achieves high-rank adaptation with improved efficiency and performance over existing low-rank methods like LoRA.

Contribution

C3A provides a high-rank, efficient alternative to LoRA, enhancing fine-tuning performance while reducing computational and memory costs.

Findings

C3A outperforms LoRA across multiple tasks.

C3A achieves higher adaptation rank with better efficiency.

Extensive experiments validate C3A's superior performance.

Abstract

Low-Rank Adaptation (LoRA) has gained popularity for fine-tuning large foundation models, leveraging low-rank matrices $\mathbf{A}$ and $\mathbf{B}$ to represent weight changes (i.e., $\Delta \mathbf{W} = \mathbf{B} \mathbf{A}$). This method reduces trainable parameters and mitigates heavy memory consumption associated with full delta matrices by sequentially multiplying $\mathbf{A}$ and $\mathbf{B}$ with the activation. Despite its success, the intrinsic low-rank characteristic may limit its performance. Although several variants have been proposed to address this issue, they often overlook the crucial computational and memory efficiency brought by LoRA. In this paper, we propose Circular Convolution Adaptation (C$^3$A), which not only achieves high-rank adaptation with enhanced performance but also excels in both computational power and memory utilization. Extensive experiments demonstrate that C$^3$A consistently outperforms LoRA and its variants across various fine-tuning tasks.