Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition
Table 1
Baseline architectures. Models A and E are two baseline architectures [13] with slightly different number of filters in some layers. Others are architectures with large-scale pooling layers or MAP layers. Conv3 indicates that the filter kernel sizes are 3×3. The ReLU gate is not shown for simplicity. The structures between small-size datasets (CIFAR, SVHN) and large-scale datasets (ImageNet) are different in input layers and fully connected layers. Parameters on the right of the slash belong to large-scale datasets.