Table 1.

Comparison of accuracy for all tasks.

Tasks	Networks	Method	Binarization	w-Reduction	Accuracy (%)
MNIST	2-layer DNN		Full precision		98.2 (0.1)
		Ours	Weights only	Manual	97.2 (0.1)
		Ours	Activation only	Manual	97.1 (0.1)
		Ours	Weights and activations	Manual	96.0 (0.2)
Dogs vs Cats	4-layer CNN		Full precision		89.7 (0.2)
		Ours	Weights only	V1	86.0 (0.3)
		Ours	Activation only	V1	84.6 (0.1)
		Ours	Weights and activations	V1	85.5 (0.6)
Spoken numbers	4-layer CNN		Full precision		93.7 (0.2)
		Ours	Weights only	V1	91.5 (0.5)
		Ours	Activation only	V1	91.4 (0.8)
		Ours	Weights and activations	V1	93.0 (0.2)
CIFAR-10	ResNet-20		Full precision		92.412
		Ours	Weights only	V1	90.2 (0.1)
		DoReFa14	Weights only		90.0
		LQ-Net12	Weights only		90.1
		DSQ13	Weights only		90.2
		ProxQuant16	Weights only		90.7
		Ours	Activation only	V2	86.2 (0.3)
		Ours	Weights and activations	V2	84.1 (0.2)
		DoReFa14	Weights and activations		79.9
		DSQ13	Weights and activations		84.1
		CL-BCNN42	Weights and activations		91.1
	VGG-Small		Full precision		93.812
		Ours	Weights only	V1	93.3 (0.1)
		BWN11	Weights only		90.1
		BinaryConnect9	Weights only		91.7
		LQ-Net12	Weights only		93.5
		Ours	Activation only	V1	92.4 (0.2)
		Ours	Weights and activations	V1	90.7 (0.2)
		XNOR-Net11	Weights and activations		89.8
		BNN10	Weights and activations		89.9
		DSQ13	Weights and activations		91.7
		CL-BCNN42	Weights and activations		92.5

The classes in all datasets are completely balanced.