Quantized Neural Networks: Training Neural Networks with Low Precision Weights and Activations