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

Table 2

Test error comparison on multiple datasets.

Method	Weight (bit)	Activation (bit)	MNIST	SVHN	CIFAR10
BNN	1	1	1.27	2.53	8.46
BWN	1	32	0.54	—	7.25
TWN	2	32	0.65	—	7.44
DoReFa	8	8	—	2.30	—
Ours	3	3	0.96	2.14	7.48