Phased pruning in neural networks recapitulates selectivity-fragility trade-offs in brain development.

Synaptic pruning refines neural circuits during development, but altered trajectories are implicated in neurodevelopmental disorders. Here, we use a task-gated neural network to show that the timing and quantity of pruning critically determine functional outcomes. Networks subjected to aggressive pruning only after an initial period of high connectivity develop enhanced resistance to interfering inputs, yet exhibit marked fragility to internal noise. In contrast, moderate pruning preserves robustness but allows greater interference. Cue-utilization diagnostics showed that apparent ambiguity tolerance in the sparsest networks was largely artifactual: at 10% early density, the network partly ignored the task cue, whereas from 20% density onward cue-dependent switching was intact. A continuous late-pruning sweep further showed that the selectivity benefit of aggressive pruning was conditional and non-monotonic, emerging most reliably when early density was high. These trade-offs emerge prominently when late pruning follows early overgrowth, providing a buffer for refinement into highly selective circuits. The findings reveal developmental phasing as a key driver of circuit specialization versus resilience, offering a computational account of one ASD-relevant trajectory in which focused selectivity coexists with fragility to perturbation, while also highlighting principles relevant to efficient sparse neural networks.