48

Algorithms for Binary Neural Networks

FIGURE 3.7

Accuracy with different θ and λ.

TABLE 3.1

Classification accuracy (%) on CIFAR-10 with 20-layer U-MCNs and MCNs.

Method

Kernel Stage

Size (MB)

WRNs

U-MCNs

MCNs

MCNs-1

MCNs

16-16-32-64

1.1

92.31

93.69

92.08

92.10

16-32-64-128

4.3

–

94.88

93.98

93.94

32-64-128-256

17.1

–

95.50

95.13

95.33

64-64-128-256

17.2

95.75

95.72

95.30

95.34

LBCNN (q=384)

–

61

–

–

92.96

–

respectively, with similar accuracy (93.98% vs. 92.96%). When LBCNN has several param-

eters (4.3M) similar to those of the MCNs, the test run time of LBCNN becomes 16.2 s,

which is still slower than our MCNs.

Visualization: We visualize MCconv features in Fig. 3.8 across different layers and the

curves of elements in different M-Filters in Fig. 3.11. Similarly to conventional CNNs,

the features of different layers capture rich and hierarchy information in Fig. 3.8. Based

on the reconstructed filters Q corresponding to the M-Filters, we obtain convolutional fea-

tures that appear diverse for different M-Filters. In summary, different MCconv layers and

Input

MCconv 1

MCconv 2

MCconv 3

FIGURE 3.8

Example of output feature maps produced by Q from different layers.