Universal pooling – A new pooling method for convolutional neural networks

Pooling is one of the key elements in a convolutional neural network. It reduces feature map size, thereby enabling training with limited amount computation. The most common pooling methods are average pooling, max and stride pooling. methods, however, have disadvantage that they can perform only specified fixed functions thus expressive power. In this paper, we propose new method named universal (UP). UP performs different depending on samples. general includes previous as special cases. structure inspired by attention methods. actually be considered channel-wise local spatial module. quite from attention-based reduction We insert into couple popular networks apply to benchmark sets two applications, namely, image recognition semantic segmentation. experiment results show complex poolings trained proposed achieves better performance than