Eidetic Learning: an Efficient and Provable Solution to Catastrophic Forgetting
Keywords: catastrophic forgetting, continual learning
TL;DR: We present a provable, scalable method for solving catastrophic forgetting.
Abstract: Catastrophic forgetting -- the phenomenon of a neural network learning a task and losing the ability to perform it after being trained on some other task -- is a long-standing problem for neural networks \citep{mccloskey1989catastrophic}. We introduce Eidetic Learning and prove that it guarantees networks do not forget. When training an EideticNet, accuracy on previous tasks is preserved because the neurons important for them are fixed and, most importantly, the hidden states that those neurons operate on are guaranteed to be unchanged by \textit{any} subsequent tasks for \textit{any} input sample. EideticNets are easy to implement, their complexity in time and space is linear in the number of parameters, and their guarantees hold for normalization layers during pre-training and fine-tuning. We show empirically with a variety of network architectures and sets of tasks that EideticNets are immune to forgetting. While the practical benefits of EideticNets are substantial, we believe they can be of benefit to practitioners and theorists alike. They have the potential to open new directions of exploration for lifelong and continual learning. We will release the code repository containing the EideticNet PyTorch framework upon publication.
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: 10935
Loading