Convolutional Neural Networks

In the previous chapters, we have seen how to construct neural networks using fully-connected layers. We will now look at a different class of layers, called convolution layers [36, 59], which are very useful when handling inputs which are images. These are routinely used in network architectures designed for tasks such as classification of images into different categories [87, 98], performing semantic segmentation on images [66, 81, 108], and transforming images from one type to another [91].