Incremental Unified Parameter Additional Tuning with Basic Memory Replaying

Jieren Deng, Jianhua Hu, Haojian Zhang, Yunkuan WangOpenReview 2022 · OpenReview

Abstract

Class incremental learning (CIL) aims to develop an open intelligence system that can continuously learn new concepts from new tasks while retaining the knowledge to distinguish between new and old concepts. Recently, parameter-additional-tuning methods (PAT) have successfully alleviated catastrophic forgetting by starting from a well-pre-trained model and only allowing a few additional parameters to be trained. However, the contradiction between stability and plasticity and the lack of inter-task features still challenge PAT-based CIL methods. To address these, we propose Unified PAT and basic memory replaying (BMR). On the one hand, unified PAT transfer the model to sequential arrival downstream tasks based on a fixed pre-trained vision transformer by unifying the prompt-based and the adapter-based methods, offering more diversified plastic structures to efficiently capture more useful features without large-scale parameters. On the other hand, BMR synthesizes on-call virtual old samples with a fixed-size basic memory to create a global task that covers up all the sub-tasks, which makes inter-task features more learnable without a large memory budget. Abundant experiments prove the effectiveness of our method.
BibTeX
@misc{deng2022incremental,
  author = {Jieren Deng and Jianhua Hu and Haojian Zhang and Yunkuan Wang},
  title = {{Incremental Unified Parameter Additional Tuning with Basic Memory Replaying}},
  year = {2022},
  url = {https://openreview.net/forum?id=CS7dB_FnGx},
  howpublished = {OpenReview},
}

Loading related content...

Figure