Pre-Defined Sparse Neural Networks with Hardware Acceleration