SSH: Sparse Spectrum Adaptation via Discrete Hartley Transformation

TL;DR

This paper introduces SSH, a spectral adaptation method using Discrete Hartley Transformation, which reduces trainable parameters and computational costs while improving performance in large model fine-tuning.

Contribution

SSH is a novel spectral adaptation approach that selects informative spectral components to enhance parameter efficiency and model performance during fine-tuning.

Findings

Outperforms existing PEFT methods in various tasks.

Reduces computational cost and memory usage significantly.

Effective in both language and multi-modality tasks.

Abstract

Low-rank adaptation (LoRA) has been demonstrated effective in reducing the trainable parameter number when fine-tuning a large foundation model (LLM). However, it still encounters computational and memory challenges when scaling to larger models or addressing more complex task adaptation. In this work, we introduce Sparse Spectrum Adaptation via Discrete Hartley Transformation (SSH), a novel approach that significantly reduces the number of trainable parameters while enhancing model performance. It selects the most informative spectral components across all layers, under the guidance of the initial weights after a discrete Hartley transformation (DHT). The lightweight inverse DHT then projects the spectrum back into the spatial domain for updates. Extensive experiments across both single-modality tasks such as language understanding and generation and multi-modality tasks such as video-text understanding demonstrate that SSH outperforms existing parameter-efficient fine-tuning (PEFT) methods while achieving substantial reductions in computational cost and memory requirements.