New Metaplasticity Rule for Continual Learning: Mitigating Catastrophic Forgetting Without Task Boundaries

In this article, the authors explore the problem of catastrophic forgetting in neural networks, which occurs when a model learns new tasks and forgets previous ones. They propose a Bayesian approach to overcome this issue, which involves incorporating uncertainty into the synaptic plasticity mechanism. The authors suggest that this approach can be analogous to synaptic local learning rate in metaplasticity, similar to how weight uncertainty value can be compared to synaptic local learning rate.
The article begins by providing context on the issue of catastrophic forgetting and how it affects neural networks. The authors explain that the human brain avoids this problem through mechanisms not fully understood but believed to involve synaptic metaplasticity, where synapses adapt their learning rate continuously rather than at the end of tasks. They then introduce their proposed approach, which is based on Bayesian inference and involves incorporating uncertainty into the synaptic plasticity mechanism.
To explain this concept further, the authors use an analogy of a car’s odometer. Just as an odometer measures the distance traveled by a car, the authors suggest that a Bayesian approach can measure the uncertainty in the synaptic plasticity process. They also provide an illustration of how this update rule works in practice, using a contour plot to visualize the density of the mean-field Gaussians for the new and previous tasks.
Throughout the article, the authors emphasize the importance of considering uncertainty in the synaptic plasticity mechanism and how it can help overcome catastrophic forgetting. They also highlight the connection between their proposed approach and recent work on Bayesian inference in neural networks.
In summary, the article presents a Bayesian approach to overcome catastrophic forgetting in neural networks by incorporating uncertainty into the synaptic plasticity mechanism. The authors use an analogy of a car’s odometer to explain this concept and provide an illustration of how the update rule works. They emphasize the importance of considering uncertainty in the synaptic plasticity mechanism and highlight the connection between their proposed approach and recent work on Bayesian inference in neural networks.