DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning

Abstract

Continual learning aims to enable a single model to learn a sequence of tasks without catastrophic forgetting. Top-performing methods usually require a rehearsal buffer to store past pristine examples for experience replay, which, however, limits their practical value due to privacy and memory constraints. In this work, we present a simple yet effective framework, DualPrompt, which learns a tiny set of parameters, called prompts, to properly instruct a pre-trained model to learn tasks arriving sequentially without buffering past examples. DualPrompt presents a novel approach to attach complementary prompts to the pre-trained backbone, and then formulates the objective as learning task-invariant and task-specific "instructions". With extensive experimental validation, DualPrompt consistently sets state-of-the-art performance under the challenging class-incremental setting. In particular, DualPrompt outperforms recent advanced continual learning methods with relatively large buffer sizes. We also introduce a more challenging benchmark, Split ImageNet-R, to help generalize rehearsal-free continual learning research. Source code is available at https://github.com/google-research/l2p.

Continual Learning for Predictive Maintenance: Overview and Challenges

“…It has been shown that catastrophic forgetting can be greatly mitigated by enriching the input patterns while using considerably less memory. In DualPrompt [212], an extension was proposed, where the authors demonstrate how prompts should be distinguished between task-specific and task-invariant.…”

Benchmarking Deepart Detection

“…Therefore, we suggest the use of the prompt-tuning techniqueJia et al (2022);Wang et al (2022b), as we empirically find it is capable of addressing the feature bias issue while it has shown a superior efficiency for transfer learningWang et al (2022c;b). (2017)) and one state-of-the-art method (S-PromptsWang et al (2022b)) as examples to illustrate the proposed framework.…”

“…To fill the gap for these models, additionally acquiring other pure conarts is nontrivial and expensive, since the Internet is now full of deeparts. An alternative solution is to reuse the used conarts of StableDiff, but explicitly accessing the previously/lately learned data is not allowed for continual learning Shokri & Shmatikov (2015); Wang et al (2022c). Therefore, we suggest keeping the natural setup where most of deepart models have no officially released conarts.…”

SLCA: Slow Learner with Classifier Alignment for Continual Learning on a Pre-trained Model

“…Continual learning aims to learn effectively from sequentially arrived data, behaving as if they were observed simultaneously. Current efforts are mainly based on the premise of learning from scratch, attempting to mitigate Classifier Alignment (SLCA) enables sequential fine-tuning (Seq FT) to outperform prompt-based approaches such as L2P [43] and DualPrompt [42] by a large margin.…”

“…Although (2) is becoming dominant in natural language processing (NLP) [17], the choice of ( 1) and ( 2) remains an open question for continual learning in computer vision (CV). The recently proposed promptbased approaches, such as L2P [43] and DualPrompt [42], followed the second strategy and reported to be far superior to the traditional continual learning baselines of finetuning the representation layer. On the other hand, since the large amount of pre-training data are typically unlabeled and may also arrive incrementally, it seems more reasonable to use self-supervised pre-training than supervised pre-training [5,40], also regarding that (upstream) continual learning in a self-supervised manner is generally more robust to catastrophic forgetting [16,25,8].…”