Squeeze-and-Threshold Based Quantization for Low-Precision Neural Networks