Deep learning using heterogeneous feature maps for maxout networks

We propose a novel type of maxout that uses filters with kernels of multiple sizes for generating convolved maps. These filters extract the most effective features for recognition from many different variations of texture patterns. A convolved map is generated by convolution between feature maps and filters. If the size of filters is varied, the size of the convolved map will also vary; in which case, since there are no correspondences among the positions of convolved maps, maxout will not work for these kinds of filters. To align the sizes of convolved maps, we converted, in advance, feature maps, which we term `heterogeneous feature maps,´ using zero padding. Converting the size of feature maps in this way allows maxout to function, even with filters of different sizes. In this study we demonstrate the classification performances using our proposed maxout on MNIST, CIFAR-10, CIFAR-100, SVHN datasets, and show a 13.17% improvement of accuracy with augmented data.