A Novel Low-Bit Quantization Strategy for Compressing Deep Neural Networks

Table 1

Expect error of different values.

Scheme	= 0.0625	= 0.125	= 0.25	= 0.5	= 1
	0.4078	0.3298	0.2106	0.0825	0.0458
	0.3298	0.2103	0.0795	0.0239	0.0443
	0.2102	0.0791	0.0209	0.0223	0.0443
	0.0790	0.0205	0.0193	0.0223	0.0443
	0.0204	0.0189	0.0193	0.0223	0.0443
	0.0189	0.0189	0.0193	0.0223	0.0443