NEAT: Nonlinear Parameter-efficient Adaptation of Pre-trained Models

Fine-tuning pre-trained models is crucial for adapting large models to downstream tasks, often delivering state-of-the-art performance. However, fine-tuning all model parameters is resource-intensive and laborious, leading to the emergence of parameter-efficient fine-tuning (PEFT) methods. One widely adopted PEFT technique, Low-Rank Adaptation (LoRA), freezes the pre-trained model weights and introduces two low-rank matrices whose ranks are significantly smaller than the dimensions of the original weight matrices. This enables efficient fine-tuning by adjusting only a small number of parameters. Despite its efficiency, LoRA approximates weight updates using low-rank decomposition, which struggles to capture complex, non-linear components and efficient optimization trajectories. As a result, LoRA-based methods often exhibit a significant performance gap compared to full fine-tuning. Closing this gap requires higher ranks, which increases the number of parameters. To address these limitations, we propose a nonlinear parameter-efficient adaptation method (NEAT). NEAT introduces a lightweight neural network that takes pre-trained weights as input and learns a nonlinear transformation to approximate cumulative weight updates. These updates can be interpreted as functions of the corresponding pre-trained weights. The nonlinear approximation directly models the cumulative updates, effectively capturing complex and non-linear structures in the weight updates. Our theoretical analysis demonstrates taht NEAT can be more efficient than LoRA while having equal or greater expressivity. Extensive evaluations across four benchmarks and over twenty datasets demonstrate that NEAT significantly outperforms baselines in both vision and text tasks.

Pear: Pruning and Sharing Adapters in Visual Parameter-Efficient Fine-Tuning

Adapters have been widely explored to alleviate computational and storage costs when fine-tuning pretrained foundation models. However, the adapter itself can exhibit redundancy, leading to unnecessary storage overhead and inferior performance. In this paper, we propose Prune and Share (Pear), a novel adapter-pruning framework for efficient fine-tuning of pretrained visual foundation models. Specifically, we prune certain adapters and share the more important unpruned ones with positions where adapters are pruned, allowing continual adaptation at these positions after pruning. Additionally, a knowledge checkpoint strategy is introduced, which preserves the information of the pruned adapters and further boosts performance. Experimental results on visual adaptation benchmark validate the effectiveness and efficiency of the proposed Pear comparing to other competitive methods. Code is in https://github.com/yibozhong/pear.

Low-Rank Interconnected Adaptation Across Layers

Low-rank adaptation (LoRA), as one of the most well-known representative methods of parameter-efficient fine-tuning, freezes the backbone model and introduces parallel adapter modules to each layer of the model. These modules consist of two low-rank trainable matrices: a low-dimension projector (LP) and a high-dimension projector (HP) with their product approximating the change for updating the model weight. However, LoRA's paired LP and HP per layer limit learned weights to specific features, ignoring the varied information extracted by stacked layers in models like Transformers. By considering the differences between layers and establishing connections across them when learning the weights, we enhance the capture of relevant information for downstream tasks using this interconnected adaptation when fine-tuning. Meanwhile, preserving the unique characteristics of each layer and thus selectively mix the learning traits of various layers according to a specific ratio can also be crucial in certain tasks. In this paper, we propose Low-rank Interconnected adaptation across layers (Lily). Specifically, we retain layer-specific LPs (local LPs) for low-dimensional feature projection and unify all HPs into a model-wide global HP, thereby overcoming layer-specific constraints in LoRA. The global HP, layer-independent, supports multiple HP sub-modules, or inspired by Mixture of Experts (MoE), HP experts capturing learning traits across all layer depths. For the ratio to mix all the experts, we use a router inspired by MoE to selectively adapt the features of different layers, thus obtaining a unique expert distribution. We evaluated Lily on a wide range of downstream tasks and achieved state-of-the-art results, outperforming LoRA and a range of competitive methods. Code will be available at https://github.com/blameitonme1/lily.

HEAT: Head-level Parameter Efficient Adaptation of Vision Transformers with Taylor-expansion Importance Scores

Prior computer vision research extensively explores adapting pre-trained vision transformers (ViT) to downstream tasks. However, the substantial number of parameters requiring adaptation has led to a focus on Parameter Efficient Transfer Learning (PETL) as an approach to efficiently adapt large pre-trained models by training only a subset of parameters, achieving both parameter and storage efficiency. Although the significantly reduced parameters have shown promising performance under transfer learning scenarios, the structural redundancy inherent in the model still leaves room for improvement, which warrants further investigation. In this paper, we propose Head-level Efficient Adaptation with Taylor-expansion importance score (HEAT): a simple method that efficiently fine-tuning ViTs at head levels. In particular, the first-order Taylor expansion is employed to calculate each head's importance score, termed Taylor-expansion Importance Score (TIS), indicating its contribution to specific tasks. Additionally, three strategies for calculating TIS have been employed to maximize the effectiveness of TIS. These strategies calculate TIS from different perspectives, reflecting varying contributions of parameters. Besides ViT, HEAT has also been applied to hierarchical transformers such as Swin Transformer, demonstrating its versatility across different transformer architectures. Through extensive experiments, HEAT has demonstrated superior performance over state-of-the-art PETL methods on the VTAB-1K benchmark.