Multiactivation Pooling Method in Convolutional Neural Networks for Image Recognition

Figure 2

Plain networks for CIFAR-10. Left: applying VGG-11 ‎[13] model, which is slightly different from the model designed for ImageNet. Middle: applying large-scale pooling (using 8×8 max-pooling to extract lower level features and 4×4 average-pooling for high-level features ‎[36]) in VGG-11 model without increasing its depth and parameters. Right: applying MAP method in the network.