Do Artificial Neural Networks Need Dendritic Nanotubes? Two Reasons for Shrinking Importance Estimates in Continual Learning
The research investigated using random links between neurons to implement importance shrinkage during continual learning. Random links were found to implement linear shrinkage toward the layer mean, improving accuracy on tasks like Permuted MNIST. This suggests that mass-conserving exchanges of plasticity signals between randomly paired units act as a statistical correction without needing explicit graph structures.