Dual Low-Rank Adaptation for Continual Learning with Pre-Trained Models

In the era of foundation models, we revisit continual learning~(CL), which aims to enable vision transformers (ViTs) to learn new tasks over time. However, as the scale of these models increases, catastrophic forgetting remains a persistent challenge, particularly in the presence of significant domain shifts across tasks. Recent studies highlight a crossover between CL techniques and parameter-efficient fine-tuning (PEFT), which focuses on fine-tuning only a small set of trainable parameters to adapt to downstream tasks, such as low-rank adaptation (LoRA). While LoRA achieves faster convergence and requires fewer trainable parameters, it has seldom been explored in the context of continual learning. To address this gap, we propose a novel PEFT-CL method called Dual Low-Rank Adaptation (DualLoRA), which introduces both an orthogonal LoRA adapter and a residual LoRA adapter parallel to pre-trained weights in each layer. These components are orchestrated by a dynamic memory mechanism to strike a balance between stability and plasticity. The orthogonal LoRA adapter's parameters are updated in an orthogonal subspace of previous tasks to mitigate catastrophic forgetting, while the residual LoRA adapter's parameters are updated in the residual subspace spanned by task-specific bases without interaction across tasks, offering complementary capabilities for fine-tuning new tasks. On ViT-based models, we demonstrate that DualLoRA offers significant advantages in accuracy, inference speed, and memory efficiency over existing CL methods across multiple benchmarks.