Adaptive Continual Learning Through Proactive Detection of Transfer and Interference

Di Shang, Man Yao, Shiyu Hu, Kexin Wang, Jiahong Zhang, bo xu, Guoqi Li

Keywords: Continual learning, lightweight finetuning

Abstract: Continual learning (CL) requires models to sequentially learn multiple tasks, maximizing transfer and minimizing interference. CL methods based on pre-trained models (PTM) have shown strong performance by integrating PTM fine-tuning with traditional approaches. Despite these promising results, current methods lack the ability to proactively detect task transfer and interference at the local optimization level, limiting their effectiveness in maximizing transfer and minimizing interference. To address this issue, we propose adaptive continual learning strategies through proactive detection of transfer and interference. We derive the conditions under which task transfer and interference occur from a model optimization perspective, based on the Fisher matrix and gradient update directions. Based on them, we proposed a task transfer distance metric to help model modules detect transfer and interference during continual learning. We propose a dynamic parameter update mechanism and a dynamic expansion strategy, based on LoRA fine-tuning and a Mixture of Experts (MoE) mechanism, to handle varying levels of task transfer and interference. Experiments results of seven benchmarks show that our method achieves the best accuracy with a limited number of parameters, maximizing transfer and minimizing interference.

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.

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: 6291