Optimize neural-network configurations across hidden-layer budgets to improve model quality under bounded compute.