Exploiting Task Relationships for Continual Learning with Transferability-aware Task Embedding
Keywords: Continual Learning, Hypernetworks, Task Embedding, Transferability
Abstract: Continual learning (CL) has been a crucial topic in contemporary deep neural network usages, where catastrophic forgetting (CF) can impede a model's ability to progressively acquire knowledge, leading to critical training inefficiency and constraint in the improvement of model's overall capacity. Existing CL strategies mostly mitigate CF either by regularizing model weights and outputs during finetuning or by distinguishing task-specific and task-sharing model components to adapt the training process accordingly. Yet despite their effectiveness, these previous explorations are mainly limited to elements of task models, while we speculate a deeper exploitation of interrelationship among tasks can provide more enhancement for CL. Therefore, to better capture and utilize the task relations, we propose a transferability task embedding guided hypernet for continual learning. By introducing the information theoretical transferability based task embedding named H-embedding and incorporating it in a hypernetwork, we establish an online framework capable of capturing the statistical relations among the CL tasks and leveraging these knowledge for deriving task-conditioned model weights. The framework is also characterized by notable practicality, in that it only requires storing a low dimensional task embedding for each task, and can be efficiently trained in an end-to-end way. Extensive evaluations and experimental analyses on datasets including Permuted MNIST, Cifar10/100 and ImageNet-R showcase that our framework performs prominently compared to various baseline methods, as well as displays great potential in obtaining intrinsic task relationships.
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: 1570
Loading