When considering the compression of neural networks, the adoption of low-bit representations for both parameters and activations has demonstrated significant efficacy. The process of learning ...