MCN: Modulated Convolutional Network

47

FIGURE 3.5

Accuracy with different K for 20-layer MCNs with width 16-16-32-64 on CIFAR-10.

columns show the accuracies of U-MCNs and MCNs, respectively. The performance in the

last three columns shows that the accuracy of MCNs only decreases slightly when binarized

filters are used. Note that with a fixed number of convolutional layers, the performance of

MCNs increases with larger network width. At the same time, the number of parameters

also increases. Compared to LBCNN, the parameters of the MCNs are much fewer (61 M

vs. 17.2 M), but the performance of the MCNs is much better (92.96% vs. 95.30%). Also,

the last three columns show that MCNs have achieved performance similar to U-MCNs and

WRNs.

3.4.5

Model Effect

Learning convergence: The MCNs model is based on a binarized process implemented

on the Torch platform (classification). For a 20-layer MCN with width 16-16-32-64 that is

trained after 200 epochs, the training process takes about 3 hours with two 1080ti GPUs. We

plot the training and testing accuracy of MCNs and U-MCNs in Fig. 3.10. The architecture

of U-MCNs is the same as that of MCNs. Figure 3.10 clearly shows that MCNs (the blue

curves) converge at speeds similar to those of their unbinarized counterpart (the red curves).

Runtime analysis: We performed a run-time analysis to compare MCNs and LBCNN.

The runtimes of MCNs and LBCNN for all CIFAR-10 test samples are 8.7 s and 160.6 s,

Conv

3×3, 80

R+

MP

Output

Input

image

B

N

Input

image

Copy

4

MP

CNN

MCN

MP: Max Pooling

R: ReLU

BN: BatchNormlization

D: Dropout

MCcov

4×3×3, 20

R+

MP

Conv

3×3, 160

R+

MP

Conv

3×3, 320

R+

MP

Conv

3×3, 640

R+

MP

FC

1024

D

B

N

MCcov

4×3×3, 40

R+

MP

B

N

MCcov

4×3×3, 80

R+

MP

B

N

MCcov

4×3×3, 160

R+

MP

FC

1024

D

Output

FIGURE 3.6

Network architectures of CNNs and MCNs.