On training simple neural networks and small-weight neurons