ABSTRACT
This chapter discusses multilayer neural networks (MLNNs) that employ convolutional and pooling layers and their combinations for constructing different deep networks that have been used to produce good performance on many long-standing computer vision problems in recent years. The sparsity that is so imposed is random and the imposition has a structure. This sparsity can be imposed in such a manner that every neuron has incoming weights from only adjacent inputs, adjacency being the important criterion. The connectivity is therefore local and sparse at the same time. Each neuron learns a feature that is local to those it connects to and learns a template that can be used only at those locations. Neural networks are not mere learning systems that learn concepts but also systems that break down data and represent data using smaller pieces of information. Using these representations, the neural network maps data to the label space.
