Convolutional Neural Networks
Feed-forward neural networks are general and versatile. But its unconstrained nature make a purely feed-forward network difficult to train and generalize, especially for datasets with prior structure. In particular, we consider data that can be represented as a grid structure such that nearby elements have strong local dependency. Common examples are images, sound, or similar sequential data with spatial orientation as key attribute.
In this chapter, we introduce the convolution operation, which can be thought of as a filter that is applied spatially in a homogenous manner. Stacking convolutional layers allow the network to learn hierarchical patterns that generalize well to test data. We will apply this architecture to image and text classification.
