Progressive Plasticity Loss Under Dale-Constrained Continual Learning

Biological nervous systems often become less plastic as experience accumulates. We ask whether an artificial network can develop a training-dependent loss of plasticity while its architecture stays fixed. We study continual learning under Dale's Principle, which constrains each neuron to have either excitatory or inhibitory outgoing weights. Dale-constrained networks appear to retain earlier tasks unusually well, improving backward transfer by 26.1 points. Much of that apparent benefit reflects a failure to acquire later tasks. Hidden units go inactive and effective representational rank declines, and across eight runs seven reach near-total rank collapse between tasks two and five, with a median onset at task three. Sparsity- and capacity-matched controls remain trainable, and a sign-shuffle control that preserves the same proportion of positive and negative weights avoids complete collapse, which localizes the failure to fixed outgoing-sign consistency at the neuron level. We call this process \textit{developmental closure}: a training-dependent reduction in learning capacity whose onset varies across runs and remains partly reversible. Our tests do not provide evidence for a critical period: changing the excitatory-to-inhibitory ratio does not systematically prevent collapse, and late recycling of inactive neurons can restore learning. Lost memory and lost plasticity then need separate remedies. Experience replay protects previously acquired information, whereas continual reinitialization preserves the capacity to acquire new tasks. Combining them raises final Split-CIFAR-10 accuracy from the $10\%$ chance level to $37.2\%$.