Task-Adaptation Curriculum Learning
Keywords: Task adaptation, transfer learning, curriculum learning, search algorithms
Abstract: A large distribution gap between a target task and pre-training tasks could undermine the task adaptation performance of pretrained models. When the target-task data are scarce, naive finetuning results in overfitting and forgetting. In various domains, skills can be transferred across semantically related tasks, among which the general-purposed ones often have more training data. Can we bridge the gap between a pre-trained model and a low-resource target task by leveraging data from other tasks? In this paper, we address the low-resource task adaptation challenge by a transfer learning curriculum, which finetunes a model on a curated sequence of intermediate tasks, thereby progressively bridging the gap between the pre-trained model and the target task. To this end, we formulate the task curriculum as a graph search problem and improve the efficiency of estimating transferability between tasks. Two search algorithms are studied, i.e., greedy best-first search and Monte Carlo tree search. We evaluate our approach, i.e., ``task-adaptation curriculum learning (TaCL)'' on two benchmark settings. Extensive evaluations on different target tasks demonstrate the effectiveness and advantages of TaCL on highly specific and low-resource downstream tasks.
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: 12688
Loading