Parallel evolutionary training algorithms for "hardware-friendly" neural networks