Balance Orthogonal Projection for Prompt in Continual Learning

The prompt tuning approach leverages pre-trained model knowledge in continual learning via prompt learning. This study shows that selecting appropriate basis vectors for orthogonal projection effectively mitigates forgetting by balancing stability and plasticity. The model’s learning and memory are enhanced by adapting the weights of projection basis vectors, which are adjusted based on task similarity, peculiarity, and the correlation of knowledge both within and beyond the memory space. We propose a novel method, Balance Orthogonal Projection for Prompt (BOP), which quantifies task overlap and measures divergence among basis vectors through singular value analysis in the memory space. BOP balances retention of prior knowledge (stability) with the integration of new information (plasticity). Knowledge importance is evaluated both across current and prior tasks and within the memory space itself. Experimental results on the 10/20-Split-CIFAR100 and 10-Split-ImageNet-R benchmarks demonstrate that our method significantly outperforms comparable approaches, confirming its effectiveness in both class-incremental learning (CIL) and task-incremental learning (TIL) scenarios.