Using Sparse Quantization for Efficient DNN Inference