DARS : ROBUST SPARSE FINE-TUNING WITH REGULARIZED SUBSPACE DISALIGNMENT

Recent works have identified the alignment, which measures a layerwise weightcorrelation, as a novel yet crucial mechanism for feature learning. We investigate anunderlying connection between the alignment learning and the structural fitting of anetwork to the training data span. Based on this insight, we further demonstrate thatfine-tuning on out-of-distribution (OOD) data disrupts this well-aligned structurefitted during the pre-training phase, degrading generalization performance. Toaddress this, we propose DARS, DisAlignment-Regularized Sparse fine-tuning, anovel sparse fine-tuning approach that mitigates disalignment by letting the gradientupdate to be partially constrained within the principal subspace of the pre-trainednetwork, constructed based on the in-distribution (ID) data used for its pre-training.Specifically, we define the two disjoint subsets of trainable parameters for sparsechannel unfreezing: i) a random subset and ii) a subset with higher gradient projections onto the principal subspace. The latter serves as a disalignment regularizerduring 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 strengthof subspace regularization, thereby balancing the trade-off between generalizationcapacity and strong fitting to new downstream tasks. By employing DARS, weachieved SOTA performance on various benchmarks, including commonsense andarithmetic reasoning tasks, across LLaMA-7B and LLaMA2-7B.