Revisiting Prompt-based Methods in Class Incremental Learning
Keywords: continual learning, pretrained-models, self-supervised learning
Abstract: In recent years, prompt-based methods have emerged as a promising direction for continual learning, demonstrating impressive performance across various benchmarks. These methods create learnable prompts to infer task identity, then select and integrate specific prompts into the pretrained model to generate instructed features for prediction. In this paper, we first analyze the working patterns of such method across different distribution scenarios through extensive empirical analysis. Our analysis exposes the limitations of existing methods: first, two-stage inference can make mistakes even when the first stage has already provided reliable predictions; second, enforcing identical architectures for both stages hampers performance gains. To address these issues, we incorporated a self-supervised learning objective to learn discriminative features, thereby boosting the plasticity of the model. During inference, we implemented a simple yet effective threshold filtering strategy to selectively pass data to the second stage. This approach prevents errors in the second stage when the first stage has already made reliable predictions, while also conserving computational resources. Ultimately, we explore utilizing self-supervised pretrained models as a unified task identity provider. Comparing to state-of-the-art methods, our method achieves comparable results under in-distribution scenarios and demonstrates substantial gains under out-of-distribution scenarios (e.g., up to 6.34\% and 5.15\% improvements on Split Aircrafts and Split Cars-196, respectively).
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
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: 2403
Loading