Convergence, divergence, and reconvergence in a feedforward network improves neural speed and accuracy