A Lightweight Binarized Convolutional Neural Network Model for Small Memory and Low-Cost Mobile Devices

Table 3

The accuracy performance of different methods is compared on the MNIST dataset.

Methods	Accuracy (%)	Params (M)
SOPCNN [49]	99.83	1.4
No routing needed [50]	99.87	1.5
Efficient-CapsNet [51]	99.84	0.16
BNN (Torch7) [3]	98.60	—
BNN (Theano) [3]	99.04	—
CBCNN (ours, bitwise operation)	99.32	0.48