Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition

Table 9

The four architectures are designed with fewer convolutional layers for CIFAR-10 datasets. Model A0 uses 2×2 max-pooling layer and 2×2 average-pooling layer. Model B0 adopts 8×8 MAP method. Model C0 uses 1×1 convolutional layer to replace the 3×3 convolutional layer right after MAP layer. Model D0 reduces one fully layer based on model C0. ReLU gates after each convolutional layer are not shown for simplicity.

VGG-9
A0	B0	C0	D0
Input (3232 RGB image)
conv3-64 maxpool 22	conv3-64 conv3-128	conv3-64 conv3-128	conv3-64 conv3-128
conv3-128 maxpool 22	conv3-128 conv3-256	conv3-128 conv3-256	conv3-128 conv3-256
conv3-128 maxpool 22	MAP 88	MAP 88	MAP 88
conv3-256 maxpool 22	conv3-256 conv3-512	conv1-256 conv3-256	conv1-256 conv3-256
conv3-256 conv3-512 avgpool 22	avgpool 44	avgpool 44	avgpool 44
fc1 512512		fc1 256256	fc1 256256
fc2 512512		fc2 256256	fc2 25610
fc3 51210		fc3 25610	-
Softmax-10