Prompt-based Continual Learning (PCL) addresses the challenge of learning new tasks without forgetting previously acquired knowledge, but existing methods are often computationally intensive. The paper One-stage Prompt-based Continual Learning presents a single-stage PCL framework, which harnesses token embeddings as prompt queries and cuts the computational cost significantly.

• Simplifies the PCL framework by removing the need for an extra query stage, thus saving about 50% computational resources.
• Incorporates Query-Pool Regularization (QR) loss during training to strengthen representation power.
• Achieves marginal accuracy loss (%) and computational benefit while outperforming two-stage PCL methods.
• Demonstrates potential for class-incremental continual learning benchmarks like CIFAR-100, ImageNet-R, and DomainNet.

This paper’s innovative approach facilitates the advancement of continual learning methods that are both efficient and effective, addressing a critical barrier in the deployment of AI applications that require continual updates or learning new information.