Bi-linearly weighted fractional max pooling

In this paper, we propose to extend the flexibility of the commonly used 2 × 2 non-overlapping max pooling for Convolutional Neural Network. We name it as Bi-linearly Weighted Fractional Max-Pooling. This proposed method enables max pooling operation below stride size 2, and is computed based on four bi-linearly weighted neighboring input activations. Currently, in a 2 × 2 non-overlapping max pooling operation, as spatial size is halved in both x and y directions, three-quarter of activations in the feature maps are discarded. As such reduction is too abrupt, amount of said pooling operation within a Convolutional Neural Network is very limited: further increasing the number of pooling operation results in too little activation left for subsequent operations. Using our proposed pooling method, spatial size reduction can be more gradual and can be adjusted flexibly. We applied a few combinations of our proposed pooling method into 50-layered ResNet and 19-layered VGGNet with reduced number of filters, and experimented on FGVC-Aircraft, Oxford-IIIT Pet, STL-10 and CIFAR-100 datasets. Even with reduced memory usage, our proposed methods showed reasonable improvement in classification accuracy with 50-layered ResNet. Additionally, with flexibility of our proposed pooling method, we change the reduction rate dynamically every training iteration, and our evaluation results indicated potential regularization effect.