Dual-Attention based prompt generation and catalyzing for instance-wise continual learning

Current prompt-based methods of continual learning usually select and learn task-level prompts which are set as additional inputs to a frozen pre-trained model. The strategy is limited by the erratic selection step and coarse-grained instructions when handling successive tasks. To handle the above issues, we propose a Prompt Generation and Catalyzing (PGC) method to softly generate and catalyze the prompts for each instance of each task. Specifically, the proposed method contains two modules including instance-specific prompt generation (ISPG) and residual prompt catalyzing (RPC). In contrast to the coarse-grained task-level prompts strategy, ISPG employs the layer attention mechanism to generate fine-grained instance-specific prompts based on layer-wise features and naive features. RPC further catalyzes the prompts to adapt to successive tasks based on the residual learning strategy to circumvent the needs of the selection step in conventional methods. Experimental results demonstrate that the proposed method achieves a remarkable improvement (by up to 30.8 %) compared to the SOTA methods on class, domain, task-agnostic, and cross-domain class continual tasks.¹