Generalizing Pooling Functions in Convolutional Neural Networks: Mixed, Gated, and Tree

We seek to improve deep neural networks by generalizing the pooling that play a central role in current architectures. We pursue a exploration of approaches to allow pooling to learn and to adapt to and variable patterns. The two primary directions lie in (1) learning a function via (two strategies of) combining of max and average pooling, (2) learning a pooling function in the form of a tree-structured fusion of filters that are themselves learned. In our experiments every pooling operation we explore improves performance when used in of average or max pooling. We experimentally demonstrate that the pooling operations provide a boost in invariance properties relative conventional pooling and set the state of the art on several widely adopted datasets; they are also easy to implement, and can be applied within deep neural network architectures. These benefits come with only a increase in computational overhead during training and a very modest in the number of model parameters.