Knowledge Accumulating Contrastive Prompt for Continual Learning
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Prompt learning, Continual learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Continual learning has been challenged by the issue of catastrophic forgetting (CF). Prompt-based methods have recently emerged as a promising approach to alleviate this problem, capturing the previous knowledge by the group of prompts. However, selecting an appropriate prompt during the inference stage can be challenging, and may limit the overall performance by the misaligned prompts.
In this paper, we propose a novel approach to prompt-based continual learning, which accumulates the knowledge in a single prompt, which has not been explored previously. Specifically, inspired by contrastive learning, we treat the input with the current and previous prompt as two different augmented views (i.e., positive pair). We then pull the features of the positive pairs in the embedding space to accumulate knowledge. Our experimental results demonstrate the state-of-the-art performance in continual learning even with a single prompt, highlighting the potential of this approach towards a `holistic' prompt for the model.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 1728
Loading