Comment by hedgehog

This technique has three significant advantages over popular low bit quantization: 1) it retains more accuracy, 2) it does not require calibration data, 3) it's easier to implement in hardware.