Suppressing recency bias through implicit task in task-agnostic continual adaptation for foundation language models
Keywords: continual learning, lifelong learning, transfer learning, foundation language models
TL;DR: We propose SRB, which suppresses recency bias in continual learning of foundation language models by leveraging implicit tasks and simple arithmetic operations without requiring task IDs, achieving excellent performance with low memory requirements.
Abstract: Foundation language models have significantly advanced natural language processing but face challenges such as catastrophic forgetting when adapting to dynamic environments with diverse tasks. Recently, among the continual learning (CL) methods for these models, model architecture expansion methods have been spotlighted due to the growth of parameter-efficient fine-tuning (PEFT) methods. However, these methods need to store past PEFT adapters for each task and require task identifiers (task IDs) to distinguish each task, thus limiting their applicability in task-agnostic settings. They also overlook recency bias, where models focus overly on current tasks at the expense of past knowledge.
To address these issues, we propose suppressing recency bias (SRB) by using the concept of implicit tasks. SRB assigns a fixed-size adapter to an implicit task, recursively storing historical knowledge through arithmetic operations with current adapters at every time step instead of task IDs. This arithmetic mitigates recency bias by integrating non-overlapping information between historical and current adapters.
Our approach requires only simple arithmetic operations without backpropagation, minimizing additional computation, and allocates a fixed-size adapter to the implicit task, resulting in low memory requirements. We evaluate SRB on CL benchmarks for foundational LMs. Experimental results demonstrate that SRB outperforms state-of-the-art methods, achieving superior generalization performance across various task sequences and models by effectively mitigating recency bias.
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: 4505
Loading