CAN - CONTINUOUSLY ADAPTING NETWORKS

Harikrishna Satheesh Pillai, Pakhi Banchalia

Keywords: Continual Learning, Catastrophic Forgetting, Synaptic Plasticity, Hebbian Learning, Adaptive Neural Networks

TL;DR: A new architecture using Hebbian Learning and Selective Training that enables Continual Learning

Abstract:

Catastrophic forgetting is a fundamental challenge in neural networks that prevents continuous learning, which is one of the properties essential for achieving true general artificial intelligence. When trained sequentially on multiple tasks, conventional neural networks overwrite previously learned knowledge, hindering their ability to retain and apply past experiences. However, people and other animals can learn new things continuously without forgetting them. To overcome this problem, we devised an architecture that preserves significant task-specific connections by combining selective neuron freezing with Hebbian learning principles. Hebbian learning enables the network to adaptively strengthen synaptic connections depending on parameter activation. It is inspired by the synaptic plasticity seen in brains. By preserving the most important neurons using selective neuron freezing, new tasks can be trained without changing them. Experiments conducted on standard datasets show that our model significantly reduces the risk of catastrophic forgetting, allowing the network to learn continually.

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: 7703