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

Table 2

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

Architecture	Accuracy (%)	Params (M)	Search methods
ResNeXt-8-64 + random erasing [43]	96.2 ± 0.06	34.4	Manual
ResNet-110 + random erasing [43]	95.9 ± 0.13	1.7	Manual
VGG8B [44]	95.47	7.3	Manual
DeepCaps [45]	94.46	7.2	Manual
WRN-28-10 + random erasing [43]	96.3 ± 0.03	36.5	Manual
DARTS (2nd order) + cutout + random erasing [47]	96.57	2.6	Gradient-based
Fine-tuning DARTS [48]	96.91	3.2	Gradient-based
Neupde [46]	92.40	0.4	Manual
CBCNN (ours, bitwise operation)	92.86	0.48	Gradient-based