Leveraging Highly Approximated Multipliers in DNN Inference