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

Table 4

Accuracy comparison with quantization of first or last convolutional layer.

CIFAR10/MNIST	BWN	BNN	Ours
+ First − last	92.37/99.37	91.40/98.66	92.08/98.86
+ First + last	92.21/99.41	91.30/98.52	91.96/98.55
− First + last	92.52/99.38	91.47/98.71	92.52/98.75
− First − last	92.75/99.46	91.54/98.73	92.12/99.04