When less is more: evolving large neural networks from small ones