Consistent Prompting for Rehearsal-Free Continual Learning

TL;DR

A novel prompt-based method, Consistent Prompting (CPrompt), is proposed to enhance prediction robustness and boost prompt selection accuracy and achieves state-of-the-art performance on multiple continual learning benchmarks.

Abstract

Continual learning empowers models to adapt autonomously to the ever-changing environment or data streams without forgetting old knowledge. Prompt-based approaches are built on frozen pre-trained models to learn the task-specific prompts and classifiers efficiently. Existing prompt-based methods are inconsistent between training and testing, limiting their effectiveness. Two types of inconsistency are revealed. Test predictions are made from all classifiers while training only focuses on the current task classifier without holistic alignment, leading to Classifier inconsistency. Prompt inconsistency indicates that the prompt selected during testing may not correspond to the one associated with this task during training. In this paper, we propose a novel prompt-based method, Consistent Prompting (CPrompt), for more aligned training and testing. Specifically, all existing classifiers are exposed to prompt training, resulting in classifier consistency learning. In addition, prompt consistency learning is proposed to enhance prediction robustness and boost prompt selection accuracy. Our Consistent Prompting surpasses its prompt-based counterparts and achieves state-of-the-art performance on multiple continual learning benchmarks. Detailed analysis shows that improvements come from more consistent training and testing.

Figures (4)

• Figure 1: Existing prompt-based methods are inconsistent between training and testing, and prompt inconsistency and classifier inconsistency are illustrated. $P_t$ and $C_t$ represent the prompt and the classifier of the current task t, respectively.
• Figure 2: Previous approaches typically train the current task prompt and classifier in isolation. Our Consistent Prompting leverages all existing prompts and classifiers to instruct the training of the current task prompt and classifier. Meanwhile, we suggest using multiple keys to map each task prompt, instead of relying on a single key, to adapt to the diverse nature of each task.
• Figure 3: The illustration of the proposed consistent prompting (CPrompt). CPrompt aims to align the training of prompts and classifiers with testing for more consistency. It consists of two main modules: classifier consistency learning (CCL, detailed in Section \ref{['sec:CCL']}) and prompt consistency learning (PCL, detailed in Section \ref{['sec:PCL']}).
• Figure 4: Illustrations of continual learning performance at each task. Each dot indicates the accuracy of the seen classes. The result of the upper-bound (UB) is represented by a dot on the overall classes.