Refined Max-Pooling and Unpooling Layers for Deep Convolutional Neural Networks

Abstract

The main goal of this paper is the introduction of new pooling and unpooling layers suited for deep convolutional neural networks. To this end, a new approximation of max-pooling inversion has been designed. The idea behind this approximation is also introduced in this paper. It is demonstrated on pools of size 2 x 2, with a stride of 2. The widely used technique of switches is combined with interpolation to form the new approximation. For that purpose, an unconventional expression of the switches has been used. Such an expression, allows the right placement of maxima in a reconstruction of original data, as well as interpolation of all unknown values in the reconstruction using the known maxima. The introduced inversion has been implemented into the aforementioned refined pooling and unpooling layers. Since they are suited for deep convolutional networks, behavior of the layers in the feed-forward and backpropagation passes had to be solved. In this context, the introduced conception of the switches has been further developed. Specifically, feed-forward and backpropagation switches are considered in the refined layers. One version of feed-forward and three versions of backpropagation switches have been introduced within this paper. The refined pooling and unpooling layers have been tested on a simple convolutional auto-encoder in order to verify functionality of the conception.