Dars : Robust Sparse Fine-Tuning with Regularized Subspace Disalignment

Abstract

Recent works have identified the alignment, which measures a layerwise weight correlation, as a novel yet crucial mechanism for feature learning. We investigate an underlying connection between the alignment learning and the structural fitting of a network to the training data span. Based on this insight, we further demonstrate that fine-tuning on out-of-distribution (OOD) data disrupts this well-aligned structure fitted during the pre-training phase, degrading generalization performance. To address this, we propose DARS, DisAlignment-Regularized Sparse fine-tuning, a novel sparse fine-tuning approach that mitigates disalignment by letting the gradient update to be partially constrained within the principal subspace of the pre-trained network, constructed based on the in-distribution (ID) data used for its pre-training. Specifically, we define the two disjoint subsets of trainable parameters for sparse channel unfreezing: i) a random subset and ii) a subset with higher gradient projections onto the principal subspace. The latter serves as a disalignment regularizer during fine-tuning, while the random subset ensures a minimal bias in parameter selection. By adjusting the ratio between the two subsets, we can control the strength of subspace regularization, thereby balancing the trade-off between generalization capacity and strong fitting to new downstream tasks. By employing DARS, we achieved SOTA performance on various benchmarks, including commonsense and arithmetic reasoning tasks, across LLaMA-7B and LLaMA2-7B.